貓咪智能項圈 · Cat Smart Collar
思考筆記 · 第三版Working notes · v3

貓咪智能項圈Cat Smart Collar

Rj · Chad · 2026 年 10 月 6 日更新Rj · Chad · updated October 6, 2026

本頁分兩部分：最上面是 10/6 的最新思路；往下是 Rj 與 Chad 的統合筆記（第二版），內容保持不變。This page has two parts: the latest thinking from October 6 at the top, then the combined notes from Rj and Chad (v2), unchanged. 跳到統合筆記Skip to the combined notes

最新 · 10/6 更新思路LATEST · OCT 6 THINKING

隨時提領的情緒價值：真實影片加上一隻會跟著真貓同步的數位分身Emotional value on demand: real clips plus a digital twin that mirrors the real cat

一句話In one line

飼主養貓要的是情緒價值，但這份價值過去只在回家陪伴時才拿得到。我們想讓飼主在任何時候、任何地方都能「提領」，靠的是兩樣東西：系統預先挑好的真實時刻，以及一隻跟著真貓狀態同步的數位分身。Owners keep cats for emotional value, but until now they could only collect it at home. We want owners to draw on it any time, anywhere, through two things: real moments the system has already picked, and a digital twin that mirrors what the real cat is doing.

1. 起點：從電子雞想到的事1. The starting point: what the Tamagotchi teaches

人們養貓狗這類傳統寵物，要的是情緒價值，而這種價值多半只能在陪伴時提供。電子雞當年會流行，正是因為它隨時都能提供情緒價值。Rj 的想法是：藉由這套系統，讓人可以隨時提領情緒價值，不需要一定回家才能提領。People keep cats and dogs for emotional value, which mostly comes only while they are together. The Tamagotchi caught on because it offered that value at any moment. Rj's idea: use this system so people can draw on emotional value whenever they want, without having to be home.

2. 看中的價值2. Why it matters

  • 說中飼主真正在買的東西。飼主買的是陪伴感，安全監控只是附帶的需求。It names what owners really buy. They buy companionship; safety monitoring is a secondary need.
  • 真實感是護城河。電子雞的情感是虛擬的，玩久會膩；你的貓今天真的在窗邊等過你，這是虛擬寵物給不了的。Authenticity is the moat. A Tamagotchi's feelings are virtual and wear thin; your cat really did wait by the window today, which no virtual pet can offer.
  • 跟寵物攝影機的差別變清楚。Furbo、Petcube 給的是「你想看就打開看」，但打開時多半只看到一隻在睡覺的貓；我們給的是「值得看的時刻，已經幫你挑好了」。A clearer difference from pet cameras. Furbo and Petcube let you look whenever you like, but you usually find a sleeping cat; we offer moments worth seeing, already picked for you.
  • 被需要的感覺。電子雞會餓、會叫你。我們可以用真實資料給出同樣的訊號，例如「你出門後牠在門口待了 20 分鐘」「今天牠比平常更常到你房間」。The feeling of being needed. A Tamagotchi gets hungry and calls you. Real data can send the same signal: "your cat sat by the door for 20 minutes after you left", "today it visited your room more than usual".

3. 怎麼提領：飼主想看的時候就有，推送只是輔助3. How owners draw on it: there when they want it, with push as a supplement

10/6 確認的方向是以飼主主動打開為主。所以產品要確保每次打開都有值得看的東西，而且要快。The direction confirmed on October 6 is owner-initiated viewing first. So every time an owner opens the app there must be something worth seeing, and it must load fast.

  • 第一個畫面不放即時影像。依序是：一句即時狀態（例如「正在你的枕頭上睡覺，已經睡了 1 小時 20 分」）、今天最新的好時刻、點一下才看即時畫面。即時畫面多半是在睡覺的貓，放第一個容易讓人失望。No live feed on the first screen. In order: a one-line live status ("asleep on your pillow for 1 hour 20 minutes"), today's latest highlight, then live view one tap away. The live feed is usually a sleeping cat, so leading with it disappoints.
  • 情緒存摺。系統全天把好時刻存起來，依情緒價值排序：互動、玩耍、看向鏡頭、可愛睡姿、等門。今天很安靜，就退回本週精選、睡姿合集或「一年前的今天」，確保打開不會空手而回。An emotional savings account. The system banks good moments all day, ranked by emotional value: interaction, play, looking at the camera, cute sleeping poses, waiting at the door. On a quiet day it falls back to the week's best, a sleeping-pose reel or "one year ago today", so no visit comes up empty.
  • 打開一到兩秒內就看到東西。短片與縮圖要事先在雲端處理好，因為想看的那一刻通常很短，例如等電梯或午休。Something on screen within one to two seconds. Clips and thumbnails are prepared in the cloud ahead of time, because the urge to look is usually brief, like waiting for a lift or a lunch break.
  • 主畫面與鎖定畫面小工具。顯示最新一張照片或一句狀態，瞄一眼就算提領一次，不用打開 App。Home and lock screen widgets. Show the latest photo or a status line, so a glance counts as a visit without opening the app.
  • 推送只留給有時效或安全相關的事件，例如等門與危險區；每日摘要由飼主自己選要不要開。這也調整了統合筆記中「主動推送帶來每天習慣」的說法。Push only for time-sensitive or safety events, such as waiting at the door and danger zones; a daily digest is opt-in. This adjusts the combined notes' line that pushed moments build the daily habit.

4. 新構想：數位分身4. The new idea: a digital twin

消費者買回來的第一步，就是替自己的貓建模：掃描貓咪照片，AI 生成一隻會動的分身，再填名字和一句話形容牠的性格（也可以空著）。之後隨著家中鏡頭每天的紀錄，把牠的動作與習慣累積成行為模型。App 中間是一隻擬真的分身，可以簡單互動；旁邊是你不在家時牠的精彩真實影片。The first thing a buyer does is model their cat: scan a few photos, let AI generate an animated twin, then add a name and a one-line description of its personality (optional). From then on, the home camera's daily footage builds a behaviour model of its movements and habits. The app centres on a lifelike twin you can interact with simply, alongside real clips of what the cat did while you were out.

  • 補上供給缺口。真貓大部分時間在睡，分身卻隨時都在。It fills the supply gap. The real cat sleeps most of the day; the twin is always there.
  • 開箱第一天就有驚喜。不必等感測資料累積，也解決了硬體產品常見的冷啟動問題。Delight from day one. No waiting for sensor data to build up, which also solves the usual cold-start problem for hardware.
  • 越用越像，形成轉換成本。用了一年的人換到別家等於從零開始，這可能比硬體更適合當護城河。The longer you use it, the more it resembles your cat. After a year, switching means starting from zero, which may make a better moat than the hardware.
設計原則：分身是鏡子，不是演員Design principle: the twin is a mirror, not an actor
  • 分身的狀態跟著真貓走：真貓在睡，分身也在睡；真貓剛在窗邊，分身就待在窗邊。The twin follows the real cat: if the cat sleeps, the twin sleeps; if it was just at the window, the twin sits there.
  • 動作與習慣都來自真實紀錄，例如慣用睡姿、最愛位置、作息時間，不憑空生成。Movements and habits come from real records, such as favourite sleeping poses, spots and routines, not invention.
  • 互動有限度，看得出是虛擬的：摸頭、叫名字有反應，但不讓分身說「我好想你」這類沒有根據的話。Interaction is limited and clearly virtual: it responds to a pat or its name, but never says ungrounded things like "I miss you".
  • 畫面上清楚區分真實與虛擬：中間是分身，真實影片明顯區隔。Real and virtual are clearly separated on screen: the twin in the centre, real clips set apart.

5. 數位分身可行性評估5. Digital twin feasibility

結論：技術上做得到，第一版不需要自己研發新技術。真正的難點在動畫設計，也就是讓一隻貓動得自然、看起來像自己家那隻。每個環節都列兩條路徑。Bottom line: it is technically achievable, and version one needs no new technology. The real difficulty is animation design: making a cat move naturally and look like the owner's own. Each part has two paths.

環節Part路徑 APath A路徑 BPath B第一版Version 1
外觀建模Appearance 2D 風格化：抓毛色、花紋、體型，套進預先設計的貓角色。Rive 這類工具做的互動動畫檔約 30 至 50 KB、手機每秒 60 格；Duolingo 也在用。團隊版每位設計師每月 25 美元。2D stylised: capture coat colour, pattern and build, applied to a pre-designed cat character. Interactive files from tools like Rive run about 30 to 50 KB at 60 fps on phones; Duolingo uses it. Team plan US$25 per editor per month. 3D 生成：Meshy、Tripo 等服務每個模型約 0.1 至 2 美元；貓毛質感不易做好，容易「不像」。3D generation: services such as Meshy and Tripo charge about US$0.10 to 2 per model; fur is hard to get right and may look off. A
讓牠動起來Animation 預製動作庫：自己設計睡覺、理毛、伸懶腰、走、跳、看窗外等動作，每隻分身共用、只換外觀。Meshy 的自動綁骨能辨識四足動物，主流的 Mixamo 只支援人形。Pre-built motion library: design sleep, grooming, stretching, walking, jumping and window-watching once, shared by every twin with only the look changed. Meshy's auto-rigging detects quadrupeds; Mixamo supports humanoids only. AI 生成影片：每秒約 0.036 美元（Seedance）、0.07 美元（Kling）到 0.40 美元（Veo 3.1）。是事先算好的影片，無法即時回應，較適合做每日動態卡片。AI video: about US$0.036 (Seedance), 0.07 (Kling) to 0.40 (Veo 3.1) per second. Pre-rendered, so it cannot respond live; better for daily story cards. A
行為模型Behaviour model 統計既有資料：用項圈事件與區域停留整理作息、最愛位置、活躍時段，沿用既有規劃，幾乎沒有額外成本。Use existing data: build routines, favourite spots and active hours from collar events and zone dwell, reusing the existing plan at almost no extra cost. 姿態估計：開源 SuperAnimal-Quadruped 涵蓋 45 種以上四足動物、含貓；但權重只限學術與非商業使用，商用需向洛桑聯邦理工學院洽談授權，或自行訓練。Pose estimation: the open SuperAnimal-Quadruped model covers 45+ quadruped species including cats, but its weights are for academic, non-commercial use; commercial use needs a licence from EPFL or our own model. A
跟真貓同步Mirroring the real cat 狀態對應：判斷真貓在睡、吃、喝、玩、窗邊或不在畫面，分身播放對應動作；藍牙約每秒更新，延遲可接受。State mapping: detect whether the cat is sleeping, eating, drinking, playing, at the window or out of view, and play the matching motion; Bluetooth updates about once a second, which is acceptable. 即時動作捕捉：從鏡頭抓姿態直接驅動分身；鏡頭晶片光做物件偵測就要每張 68 至 97 毫秒，只能放雲端，延遲與成本都高。Live motion capture: drive the twin from camera pose in real time; the camera chip already needs 68 to 97 ms per frame just for detection, so this means the cloud, with higher latency and cost. A
互動Interaction 預製互動：摸頭、逗貓棒、餵零食有對應反應；Rive 的狀態機正是為互動角色設計。Pre-built interactions: responses to a pat, a wand toy or a treat; Rive's state machine is built for interactive characters. AI 驅動個性：依飼主描述的性格回應；容易說出沒有根據的話，違反鏡子原則。AI personality: replies based on the owner's description; risks ungrounded lines that break the mirror principle. A
開箱建模Onboarding 開箱拍 3 至 5 張照片，填名字與一句性格（可空）；第一天就有儀式感，但照片品質不一。Take 3 to 5 photos at unboxing, add a name and an optional line on personality; a moment of ceremony on day one, though photo quality varies. 用前幾天鏡頭畫面自動建模；特徵較準，但驚喜晚幾天。Build automatically from the first days of camera footage; more accurate, but the delight comes days later. A + B

開箱建模建議 A 加 B：先用照片做第一版分身，幾天後用鏡頭資料自動修正，並通知飼主「我們更認識牠了」，這本身也是一個情緒時刻。For onboarding, A plus B: a first twin from photos, refined automatically from camera data a few days later, with a note to the owner that "we know your cat better now", which is an emotional moment in itself.

  • 最大的成本與風險在美術：需要一位懂動物動態的動畫設計師，費用與人選尚未查。The biggest cost and risk is art: we need an animator who understands animal movement; cost and candidates not yet researched.
  • 待確認：姿態估計模型的商用授權；飼主覺得分身「像不像我家的貓」，只能靠測試回答。Open items: a commercial licence for pose estimation; whether owners feel the twin looks like their cat, which only testing can answer.
  • 之後的升級：3D 擬真分身、即時動作捕捉、AI 個性，放到第二階段。Later upgrades: lifelike 3D twins, live motion capture and AI personality belong in Phase 2.

6. 市場查核6. Market check

數位分身。PetLive 在 2026 年 2 月推出：用照片生成擬真 3D 寵物分身、可互動，再搭配實體相框在家播放，主打寵物過世後的陪伴與紀念。Petpix 這類 App 只把照片轉成不會動的卡通頭像。還沒看到「分身連結真實鏡頭與感測資料、跟著真貓同步」的產品，這可能是我們的差異點。Digital twins. PetLive launched in February 2026: lifelike, interactive 3D pet avatars from photos, plus a physical frame for the home, aimed mainly at comfort and remembrance after a pet dies. Apps like Petpix only turn photos into static cartoon portraits. We have not found a twin linked to real camera and sensor data that mirrors the live cat, which may be our point of difference.

全天心情分數。目前沒有產品完全做到「全天監測、以分數即時呈現貓的心情」。Catlog 有壓力分數，但只有三級、以 30 天基準比較，連續幾天異常才提醒；PetPace 量心率變異但偏健康監測；杭州的 PettiChat 以叫聲加姿態即時判斷行為狀態，宣稱受控條件下 94.6%；MeowTalk 與 CatsMe 需要飼主主動錄音或拍照。一篇 2026 年的評測更指出，能分辨貓情緒的項圈大致上不存在，部分標榜情緒 AI 的產品只是播放預錄句子。All-day mood scores. No product yet delivers an all-day, real-time mood score for cats. Catlog has a stress score, but with only three levels against a 30-day baseline, alerting after several days; PetPace measures heart rate variability but leans toward health; Hangzhou's PettiChat reads behaviour from meows and posture in real time, claiming 94.6% under controlled conditions; MeowTalk and CatsMe need the owner to record or photograph. A 2026 review found emotion-reading cat collars largely do not exist, and some "emotion AI" products just play pre-recorded phrases.

含意：介面可以像心情分數，但內容寫成可由行為推得的指標，例如活躍度、放鬆程度、互動量，再搭配和牠平常比的變化，避免過度承諾。Implication: the interface can feel like a mood score, but its content should be behaviour-based indicators such as activity, relaxation and interaction, compared with the cat's own normal, to avoid over-promising.

7. 要小心的地方7. Watch-outs

  • AI 擬人化的界線。貓咪心聲配音適合娛樂分享；但在飼主面前持續編造沒有根據的話，會稀釋真實感這個護城河。情緒語句要對應真實行為，配音清楚標示是好玩的。The limit of AI anthropomorphism. Cat voiceovers suit fun sharing, but inventing ungrounded lines for the owner erodes the authenticity moat. Emotional lines must match real behaviour, and voiceovers are clearly labelled as playful.
  • 紀念用途。行為模型越像，寵物過世後分身就越可能成為強烈的情感需求。這可以是溫暖的服務，也可能被認為在消費悲傷。初期不主打，但要先想好貓過世時分身與資料怎麼處理。Remembrance. The closer the model, the more the twin may matter after a pet dies. That can be a warm service or be seen as profiting from grief. Do not lead with it, but decide early how the twin and data are handled when a cat dies.
  • 品牌主軸（建議，待確認）。「隨時提領情緒價值」偏溫暖，「先讓問題被看見」偏焦慮。建議以情緒價值為主軸、照護為附帶；焦慮型內容留在推下一批產品的那一刻。Brand lead (suggestion, to confirm). "Emotional value on demand" is warm; "make the problem visible" is anxious. Suggest leading with emotional value, with care as secondary, and keeping anxiety-led content for the moment we offer prevention products.

8. 對第一階段與驗證的影響8. What changes for Phase 1 and testing

  • 分身分兩步。開箱先做簡單版：照片生成風格化頭像，加上名字與一句性格，讓開箱有儀式感；行為模型在背後累積，會動、會同步的分身放在第一階段後段。The twin in two steps. Start simple at unboxing: a stylised portrait from photos, plus a name and a line on personality; the behaviour model builds quietly, and the animated, mirrored twin comes later in Phase 1.
  • 需求驗證不必先做出分身。請設計師替每個試用家庭的貓畫一張風格化角色、加幾個簡單動作示意，觀察飼主覺得像不像，以及有了分身是否更常打開。No need to build the twin to test demand. Have a designer draw a stylised character with a few simple motions for each test household's cat, then see whether owners find it a likeness and open the app more often.
  • 情緒存摺也能先手動模擬。用共享相簿每天放入挑好的片段，觀察試用家庭多常主動去看，不必寫程式。The savings account can be simulated by hand. Add curated clips to a shared album daily and see how often households look, with no code.
新增要追蹤的指標New metrics to track驗證什麼What it tests
每天主動打開幾次，其中在不在家時段的比例Daily owner-initiated opens, and the share while away from home「隨時提領」是否成立Whether on-demand value holds
打開後至少看完一段片段的比例Share of visits where at least one clip is watched內容是否值得看Whether content is worth seeing
空手而回比例（打開三秒內就關掉）Empty-visit rate (closed within three seconds)要壓低；反映存摺是否夠用Keep low; shows whether the account is stocked
飼主覺得分身像不像Whether owners see the twin as a likeness分身值不值得做Whether the twin is worth building
與統合筆記的關係（提案，待確認）How this fits the combined notes (proposal, to confirm)

Chad 的價值鏈「感測 → 事件 → 影像證據 → AI 解讀 → 情感洞察 → 行動」，說明系統背後怎麼運作；Rj 的「隨時提領」與數位分身，是飼主實際感受到的體驗。產品主張或許可以從「每天一扇看進貓咪生活的窗」，再推進一步成為「想牠的時候，牠一直都在」。Chad's value chain (sensor → event → visual evidence → AI interpretation → emotional insight → action) explains how the system works underneath; Rj's on-demand value and digital twin are what owners actually feel. The proposition could move from "a daily window into your cat's life" one step further to "whenever you miss your cat, it is there".

10/6 資料來源October 6 sources
以下為統合筆記 · 第二版（Rj 與 Chad）Below: combined notes · v2 (Rj and Chad)
思考筆記 · 第二版Working notes · v2

統合筆記：貓咪智能項圈Combined notes: Cat Smart Collar

Rj（策略與行銷）· Chad（故事線與審閱）· 2026 年 10 月 5 日Rj (strategy and marketing) · Chad (storyline and review) · October 5, 2026

本版更新What changed in v2

依 Chad 審閱意見改寫：先講結論的執行摘要（為何做、為何是現在）；故事線改為「問題 → 缺口 → 主張 → 情境 → 商業模式」，技術放在後面；新增先做貓的理由、台灣市場與嬰兒市場對比、分階段（第二階段募資做醫療）、競品成本比較、銷售通路與 16 週驗證時程。Rj 的技術細節完整保留在附錄。內容仍是假設，需與合作夥伴討論並實測。Rewritten from Chad's review: a conclusion-first executive summary (why this, why now); the storyline now runs problem → gap → proposition → use cases → commercial model, with the technology after; added why cats first, the Taiwan pet versus baby market contrast, phasing (Phase 2 raises capital for medical), competitor cost comparison, sales channels and a 16-week validation timeline. Rj's technical detail is kept in full in the appendix. Everything is still a hypothesis to discuss with partners and test.

01 · 執行摘要：為何做、為何是現在EXECUTIVE SUMMARY: WHY THIS, WHY NOW

先做貓、先做台灣：用輕量感測項圈搭配家用鏡頭，每天告訴飼主「貓今天做了什麼」。建立品牌與市佔後，再募資進行第二階段的完整醫療研發。Cats first, Taiwan first: a light sensor collar plus a home camera that tells owners what their cat did today. Once the brand and market share are built, raise capital for Phase 2: full medical R&D.

  1. 01
    為何做：市場先解決了錯的問題。Why this: the market is solving the wrong problem first.現有穿戴裝置著重「能量測什麼」，飼主還得自己打開 App、解讀數據。機會在回答「我家貓今天發生了什麼，有沒有我該知道的事？」情感給人使用的理由，技術給人相信的理由。Wearables focus on what they can measure; owners still have to open an app and interpret the data. The opening is to answer "What happened to my cat today, and is there anything I should know?" Emotion gives the reason to engage; technology gives the reason to believe.
  2. 02
    為何是台灣：貓在增加，嬰兒在減少。Why Taiwan: more cats, fewer babies.家貓約 174 萬隻、首度超過家犬，成長來自都會多貓家庭。寵物市場已超過新台幣 815 億元，三年內可望破千億；新生兒十年減半，2025 年僅 107,812 人。About 1.74 million pet cats now outnumber dogs, with growth from multi-cat urban homes. The pet market has passed NT$81.5 billion and may top NT$100 billion within three years, while births have halved in ten years to 107,812 in 2025.
  3. 03
    為何是現在、為何先做貓：好驗證，元件便宜，市場有空檔。Why now, and why cats first: easy to validate, cheap parts, an open gap.貓多半養在室內，活動範圍就是家裡，項圈加鏡頭就能涵蓋，最容易驗證；狗要外出，變數多很多，留到之後。元件也已便宜現成（項圈約 5.5 至 12 美元、AI 鏡頭模組約 40 至 60 美元）。對手各做一塊，大品牌已開始做多貓辨識，窗口不會一直開著。Cats mostly live indoors, so the home is their whole world and a collar plus a camera covers it, which makes them the easiest to validate. Dogs go outside and bring far more variables, so they come later. Parts are cheap and off the shelf (collar about US$5.50 to 12, AI camera module about US$40 to 60). Competitors each cover one piece, and large brands are moving into multi-cat identification, so the window will not stay open.
  4. 04
    為何不先做醫療：太重、太慢，放到第二階段。Why not medical first: too heavy and slow, so it is Phase 2.現在就做 Garmin 等級的完整健康與醫療數據，代表更長的研發、更重的硬體與法規負擔、更難懂的訴求，也更慢驗證。第一階段只做行為與情感，喝水、活動變化以趨勢呈現、不做診斷；成功進入市場並取得足夠市佔後，再募資投入醫療研發。Building full Garmin-style health and medical data now means longer R&D, heavier hardware and regulatory load, a harder message and slower validation. Phase 1 covers behaviour and emotion only, with drinking and activity shown as trends, never diagnoses. Once market entry succeeds and share is large enough, we raise capital for medical R&D.
  5. 05
    怎麼賺錢：模式已被市場接受。How it pays: the model is already accepted.飼主第一年已為單一功能設備支付約 119 至 370 美元、每月 5 至 14 美元訂閱。硬體帶進用戶，訂閱帶來獲利，資料累積後再賣對應的預防型產品；這些營收與市佔就是第二階段募資的依據。Owners already pay about US$119 to 370 in year one for single-purpose devices, plus US$5 to 14 a month. Hardware brings users in, subscriptions make the profit, and once data builds up we sell matching prevention products. That revenue and share are the case for the Phase 2 raise.
需要的決定Decision needed

核准約 16 週的低成本測試，找出飼主「每天都想看」的 3 至 5 個時刻，並確認技術能穩定偵測，再決定是否投入打樣。Approve a low-cost test of about 16 weeks to find the 3 to 5 moments owners want every day and confirm the technology can detect them reliably, before funding prototypes.

02

飼主以為貓在家很安全，其實問題只是沒被看見Owners assume their cat is safe at home; the problems just go unseen

So what

就算裝了攝影機，飼主拿起手機那幾秒剛好撞見危險的機率很低。目標客群正是台灣快速增加的都會多貓家庭：他們在意、願意花錢，卻看不到貓獨處時的生活。Even with a camera, the odds of catching a risky moment in the few seconds an owner checks are low. Our target is Taiwan's fast-growing multi-cat urban households: they care and they spend, but they cannot see what their cats do alone.

客群在哪：先做台灣Where the customer is: Taiwan first

1.74M家貓，兩年增 32.8%，首度超過家犬約 146 萬隻（農業部 114 年調查）pet cats, up 32.8% in two years, now above about 1.46M dogs (Ministry of Agriculture, 2025)
NT$81.5B+2025 年寵物市場規模，三年內可望破千億（台經院）pet market in 2025, expected to pass NT$100B within three years (TIER)
NT$30K+每隻犬貓每年平均花費（台經院）average yearly spend per dog or cat (TIER)
107,8122025 年新生兒，年減 20.5%，連續十年創新低，約為 2015 年的一半（內政部）births in 2025, down 20.5%, a tenth straight record low and about half of 2015 (Ministry of the Interior)

寵物在上升，嬰兒相關產業在下降Pets rising, baby-related industries declining

為什麼是台灣，而不是先做其他市場Why Taiwan rather than another market first

第二階段可評估的海外市場（建議，尚未研究）：日本、香港、新加坡等貓文化強的亞洲市場。Overseas markets to assess in Phase 2 (suggestion, not yet researched): cat-loving Asian markets such as Japan, Hong Kong and Singapore.

03

缺口：追蹤、監測、內容各有產品，「理解」沒有The gap: tracking, monitoring and content are covered; understanding is not

So what

把飼主的行為從「我得去看看貓」變成「我的貓有東西要給我看」。主動推送的時刻帶來每天的使用習慣，習慣帶來訂閱。Shift owner behaviour from "I need to check on my cat" to "my cat has something to show me". Moments pushed to the owner create a daily habit, and the habit supports a subscription.

現有產品大致落在三類：追蹤（牠在哪）、監測（牠是否活躍、健康、安全）、內容（拍到可愛畫面）。我們要把三者接成第四類：理解，也就是「我不在家時，牠到底在做什麼？」Today's products sit in three groups: tracking (where is my pet), monitoring (is it active, healthy, safe) and content (capture something cute). The opportunity is to join them into a fourth: understanding, meaning "what is my pet actually doing when I am not there?"

感測Sensor→事件Event→影像證據Visual evidence→AI 解讀AI interpretation→情感洞察Emotional insight→行動Action
層次Layer飼主得到的價值Customer value商業價值Business value
情感Emotion每天的連結與安心Daily connection and reassurance使用黏著與留存Engagement and retention
洞察Insight不在家也能理解牠的行為Understand behaviour while away資料與訂閱收入Data and recurring subscription
行動Action發現風險或未被滿足的需求Spot risks or unmet needs通往後續產品與服務Pathway to products and services

全球寵物穿戴市場 2026 年估計約 38 億美元，智慧項圈是主要型態（Grand View Research）。The global pet wearable market is estimated at about US$3.8 billion in 2026, with smart collars the dominant format (Grand View Research).

04

「每天一扇看進貓咪生活的窗」：先做貓的情感與行為，醫療研發等募資後"A daily window into your cat's life": cats, emotion and behaviour first; medical R&D after a capital raise

So what

一句話主張：「每天自動收到你家貓今天的精華。」項圈與鏡頭只是基礎設施，產品是飼主與貓的關係。先做貓，因為貓多在室內、變數少、最好驗證。現在就做 Garmin 等級的完整醫療數據太重，所以第一階段建立品牌與市佔，第二階段再以募資投入醫療研發。One-line pitch: "Every day, automatically receive your cat's highlights." The collar and camera are infrastructure; the product is the relationship. Cats come first because they live mostly indoors, with fewer variables, so they are the easiest to validate. A full Garmin-style medical spec is too heavy for now, so Phase 1 builds the brand and market share, and Phase 2 uses a capital raise to fund medical R&D.

照護價值留住用戶，創作價值帶進新用戶。定位不主打「人人都是貓咪創作者」，而是「飼主自己每天想看的內容」，分享是附帶效果。Care value keeps users; creative value brings new ones. Not "everyone can become a cat creator" but "content owners want to watch themselves", with sharing as a side effect.

第一階段 · 現在PHASE 1 · NOW先做貓：建立品牌與市佔Cats first: build the brand and share

  • 台灣室內貓，項圈加一台鏡頭；每日精華、區域報告、危險區警報Indoor cats in Taiwan; collar plus one camera; daily highlights, zone report, danger alerts
  • 盡量採購現有硬體，力氣放在整合與 App；不做醫療Buy existing hardware where possible; effort goes into integration and the app; no medical features
  • 加購：多貓家庭的額外項圈、區域標籤或房間接收器、可換項圈帶、選配項圈攝影機（冒險模式）Add-ons: extra collars for multi-cat homes, zone tags or room receivers, swappable bands, optional collar camera (adventure mode)
  • 資料累積後推出 Rj 所說的「下一批產品」，依問題對應預防型產品：窗邊陽台→防護網；跳爐台→保護蓋；喝水少→智慧飲水機；獨處不動→自動逗貓器；搶食追打→分開餵食Once data builds up, launch what Rj calls the next wave: prevention products matched to it: windows and balconies → guards; stove jumping → covers; low drinking → smart fountain; inactive alone → automatic toy; food fights → separate feeders

為什麼是現在、為什麼是貓Why now, why cats

  • 好驗證：貓多在室內，家就是牠的全部活動範圍；狗要外出，變數多很多Easy to validate: cats mostly stay indoors, so the home is their whole range; dogs go outside, with far more variables
  • 需求：家貓超越家犬、多貓家庭增加，寵物支出上升而嬰兒市場萎縮Demand: cats now outnumber dogs, multi-cat homes are rising, pet spending grows while the baby market shrinks
  • 成本：感測元件、AI 鏡頭晶片與 AI 配音都已便宜現成Cost: sensors, AI camera chips and AI voice are cheap and off the shelf
  • 空檔：還沒看到整合型產品，但 Petlibro 2026 年 9 月已推出多貓辨識餵食器Gap: no integrated product found yet, but Petlibro launched a multi-cat recognition feeder in September 2026

第二階段 · 募資之後PHASE 2 · AFTER A CAPITAL RAISE完整醫療研發Full medical R&D

  • 投入完整健康與醫療監測的研發（Garmin 等級的生理數據），可與動物醫院合作R&D for full health and medical monitoring (Garmin-style physiological data), potentially with vet clinics
  • Rj 標示為「之後」的項目：UWB 精準定位的高階版、與 Bugaboo 等品牌的聯名項圈帶Items Rj marked for later: a premium version with UWB positioning, co-branded bands with brands such as Bugaboo
  • 擴展到狗（戶外、變數較多）；海外市場待評估（建議，尚未研究）Expansion to dogs (outdoors, more variables); overseas markets to assess (suggestion, not yet researched)

募資條件（數字待與 Rj 訂定）Conditions for the raise (thresholds to set with Rj)

  • 成功進入市場：飼主每天使用、付費轉換達標Successful market entry: daily use and paid conversion meet target
  • 足夠的市佔與品牌知名度，支撐募資估值Enough market share and brand awareness to support a raise
  • 偵測準確度與信任已建立，飼主相信我們的資料Proven detection accuracy and trust: owners believe our data
05

用 3 至 5 個情感鉤子建立每天打開的習慣Build the daily habit on 3 to 5 emotional hooks

So what

技術要從飼主在乎的時刻倒推，而不是反過來。關鍵問題是：哪 3 至 5 個時刻夠感人、夠有用、技術也做得到，讓飼主說「我每天都要」？前四個是最有希望的候選，由測試決定。Build the technology backwards from the moments owners care about, not the other way round. The question: which 3 to 5 moments are emotional, useful and achievable enough that an owner says "I want this every day"? The first four are the strongest candidates; testing decides.

時刻Moment飼主看到的What the owner sees價值Value
等門計時Waiting at the door「你回家前，牠在門口等了 18 分鐘。」"Your cat waited by the door for 18 minutes before you came home."情感、分享Emotion, sharing
凌晨暴衝Night zoomies半夜三點的跑酷，自動剪成片段The 3 a.m. sprint, cut into a clip automatically分享Sharing
抓元凶Who did it「Mochi 下午 2:14 把它弄掉了。」"Mochi knocked this over at 2:14 p.m."照護、分享Care, sharing
多貓關係Multi-cat relationships誰黏誰、誰追誰、誰在躲誰；「兩隻今天在一起 47 分鐘」Who sticks to, chases or avoids whom; "your two cats spent 47 minutes together today"照護、分享Care, sharing
牠的一天A day in its life睡在哪、睡多久、動了多少Where and how long it slept, how active it was照護Care
窗邊、陽台、流理台Windows, balconies, counters在危險或禁止區域停留的次數與時間How often and how long it stayed in danger or restricted zones安全，帶動預防型產品Safety, drives prevention products
喝水與活動變化Drinking and activity changes和上週比是否明顯變少Whether either dropped sharply against last week照護，帶動預防型產品Care, drives prevention products
06

硬體 → 訂閱 → 生態系Hardware → subscription → ecosystem

So what

硬體毛利薄又容易被複製，用來帶進用戶；穩定收入來自訂閱；資料累積後在 App 內銷售預防型產品，擴大每戶營收。這些營收與市佔是第二階段募資的基礎。多貓家庭加購項圈是主要成長點，安全警報免費以保護信任。Hardware is thin-margin and easy to copy, so it acquires users; subscriptions bring steady income; once data builds up, prevention products sold in the app grow revenue per home. This revenue and share are the basis for the Phase 2 raise. Extra collars for multi-cat homes are the main growth lever, and safety alerts stay free to protect trust.

收入來源Revenue line內容What it is
入門組合Starter kit一個項圈加一台鏡頭One collar and one camera
加購Add-ons多貓家庭的額外項圈；區域標籤或房間接收器；可換項圈帶（參考 Swatch 錶帶）；選配項圈攝影機。聯名款（如 Bugaboo）放到之後Extra collars for multi-cat homes; zone tags or room receivers; swappable bands (Swatch strap model); optional collar camera. Co-branded editions (e.g. Bugaboo) come later
免費版Free tier當日精華、基礎區域統計、極危險區警報Daily highlights, basic zone stats, danger zone alerts
付費版Paid tier影片長期保存、AI 配音額度、完整每日報告、異常提醒；參考區間每月 5 至 14 美元，定價待定Long-term clip storage, AI voiceover credits, full daily reports, anomaly alerts; reference band US$5 to 14 a month, pricing open
預防型產品Prevention products依資料推薦、App 內銷售（第一階段後段）Recommended from the data and sold in the app (later in Phase 1)

銷售通路：先 KOL 與線上，再進連鎖門市Sales channels: KOLs and online first, then retail chains

分享影片本身就是獲客管道。連鎖總部進場需要試賣成績，所以先用線上與社群累積銷售數據。Shared clips are an acquisition channel in themselves. Chain headquarters want trial sales results, so build sales data online and through social first.

通路Channel角色Role時機When
KOL 與社群分享KOLs and social sharing雙方的寵物圈人脈；每支分享影片帶品牌浮水印，形成內容飛輪Partners' pet-industry network; every shared clip carries a watermark, feeding a content flywheel測試期起From testing
群眾募資（如嘖嘖）Crowdfunding (e.g. zeczec)預購同時驗證付費意願與籌措首批資金（建議，待討論）Pre-orders that test willingness to pay and fund the first batch (suggestion, to discuss)打樣後After prototype
自營官網與 LINE 官方帳號Own site and LINE official account直接掌握用戶與訂閱關係Direct ownership of users and subscriptions上市Launch
電商平台：momo、蝦皮、PChomeMarketplaces: momo, Shopee, PChome觸及量；momo 上貓用品與狗用品銷量約 6:4，貓用品成長較快Reach; on momo, cat and dog products sell about 60:40, with cat products growing faster上市Launch
寵物連鎖：東森寵物（109 家）、寵物公園（100 家以上）Pet chains: ETtoday Pets (109 stores), Pet Park (100+)實體體驗與信任；總部審核嚴謹、通常要求試賣Hands-on demos and trust; strict HQ review, usually a trial sale first有銷售數據後Once sales data exists
經銷商與獨立寵物店Distributors and independent shops新品牌常見的低成本起點A common low-cost starting route for new brands上市後After launch
動物醫院、貓旅館Vet clinics, cat hotels照護與安全訊息的可信推薦者，也是醫療研發的合作對象（待驗證）Credible advocates for care and safety, and partners for medical R&D (to validate)第二階段Phase 2
07

項圈負責感測與身分，鏡頭負責畫面，軟體把兩者變成故事The collar senses and identifies, the camera films, the software turns both into a story

So what

不把項圈做成迷你攝影機或醫療設備。項圈提供鏡頭欠缺的「哪隻貓」與「精確時間」，讓系統能自動挑出時刻。The collar does not become a mini camera or a medical device. It supplies what cameras lack, which cat and exactly when, so the system can pick moments automatically.

系統怎麼運作How the system works

定位提醒：不主打「人人都是貓咪創作者」，而是「飼主自己每天想看的內容」，分享是附帶效果。推送的問題必須是真實發生的事，誤報或誇大會讓飼主覺得被推銷。Positioning note: not "everyone can become a cat creator", but "content owners want to watch themselves", with sharing as a side effect. Anything we surface must have really happened; false or exaggerated alerts would make owners feel sold to.

08

第一階段所需功能都有可行路徑；風險在實際家庭的準確度Every Phase 1 feature has a workable path; the risk is accuracy in real homes

So what

原則是「項圈觸發、鏡頭確認」。不需要發明新技術，但實驗室數字推到不同貓與家庭會下降多少仍是未知，必須在測試階段量出來。The principle is "the collar triggers, the camera confirms". No new technology is needed, but how far lab figures drop across different cats and homes is unknown and must be measured in the test phase.

功能Capability第一階段做法Phase 1 approach依據Evidence信心Confidence
追逐、高速移動Chasing, fast motion加速度計分類，鏡頭確認Accelerometer classification, camera confirms研究在 12 種行為達約 99%，15 種約 70% 以上Studies: about 99% across 12 behaviours, above 70% across 15中高Med-high
跳躍、爬高Jumping, climbing加速度計偵測跳躍，氣壓計確認高度Accelerometer spots the jump, barometer confirms height氣壓精度約 ±25 公分：一公尺以上可測，40 公分沙發在邊緣Pressure accuracy about ±25 cm: above 1 m detectable, a 40 cm sofa is marginal中Medium
兩貓互動Two-cat interaction藍牙判斷「非常靠近」，鏡頭確認Bluetooth for "very close", camera confirms1 公尺內誤差約 10 公分，距離拉遠誤差變大About 10 cm error within 1 m, growing with distance中Medium
哪一隻貓Which cat項圈藍牙識別碼對應畫面Collar Bluetooth ID matched to footage識別碼可靠；對應準確度需實測ID is reliable; matching needs testing中高Med-high
區域與危險警報Zones and danger alerts畫面內用視覺圈選；畫面外用藍牙房間接收器；兩者交叉確認Vision for zones in view; Bluetooth room receivers out of view; cross-check both藍牙約房間級；視覺每秒 10 至 15 張足以判斷停留Bluetooth is about room-level; vision at 10 to 15 fps is enough to judge dwell中Medium
剪輯與配音Clipping and voiceover本地循環錄影，只上傳事件前後；雲端 AI 語音Local loop recording, upload only around events; cloud AI voice配音每句約 0.0002 至 0.002 美元Voiceover about US$0.0002 to 0.002 per line高High

最大待確認項：市售鏡頭能否取得影像（很多封閉在自家雲端）、氣壓是否受開門與冷氣干擾。細節見附錄 A。Biggest open items: whether retail cameras expose their video (many lock it in their own cloud), and whether doors and air conditioning disturb pressure readings. Detail in Appendix A.

主要風險：信任（誤報與資安）Key risks: trust (false alerts and privacy)

風險Risk對策Mitigation
誤報傷害信任False alerts damage trust影像與藍牙交叉確認；危險區設停留秒數門檻Cross-check vision and Bluetooth; dwell threshold for danger zones
危險警報漏報Missed danger alerts定位為輔助提醒而非安全保證，合約與 App 寫清楚Position as an assistive alert, not a safety guarantee, stated in terms and in the app
中國製鏡頭資安疑慮Privacy concerns over Chinese-made cameras資料放台灣或日本；自寫韌體較容易做到Host data in Taiwan or Japan; easier with our own firmware
實驗室精度不等於實際Lab accuracy may not carry over用真實家庭資料重新評估Re-evaluate on real household data
市售鏡頭取不到影像Retail cameras may not expose video選購前確認本地串流或開放介面Confirm local streaming or open interfaces before choosing
貓不願配戴Cats refuse to wear it盡量輕，使用安全分離扣Keep it light; breakaway buckle
法規Regulation上市前 NCC 認證；影像依個資法設計同意與刪除NCC certification before launch; consent and deletion for video under the privacy law
硬體易被複製Hardware is easy to copy優勢放在 App 體驗、通路與內容Build the edge in the app, distribution and content
社群自動發文限制Auto-posting limits on social platforms用手機分享，不承諾全自動上傳Use the share sheet; don't promise automatic upload
09

項圈元件約 5.5 至 12 美元，加鏡頭約 46 至 77 美元（未含電路板與組裝）Collar parts about US$5.50 to 12; with a camera about US$46 to 77 (before board and assembly)

So what

元件成本不是瓶頸，訂閱與 App 才是價值所在。電路板、外殼、組裝與雲端成本尚未詢價，是測試階段第 5 步要補齊的數字。但要注意，2026 年中國上游材料與記憶體正在漲價：8 月生產者物價年增 3.8%，其中電子設備製造年增 5.3%；DRAM 合約價第一季季增約九成，新產能最快 2027 年底才出來。項圈端受影響較小，需要記憶體與儲存的鏡頭端成本要保守估。Parts cost is not the constraint; the value sits in the subscription and the app. Board, housing, assembly and cloud costs are not yet quoted and are filled in at step 5 of the test phase. Note, though, that Chinese upstream materials and memory are getting more expensive in 2026: producer prices rose 3.8% year on year in August, with electronic equipment manufacturing up 5.3%; DRAM contract prices rose about 90% quarter on quarter in Q1, with new capacity not expected before late 2027. The collar is less exposed; the camera side, which needs memory and storage, should be costed conservatively.

多取自 LCSC 網路報價，實際代工報價可能不同。電路板、充電電路、外殼、項圈帶與組裝尚未詢價。Mostly LCSC online quotes; factory quotes may differ. Boards, charging, housing, bands and assembly not yet quoted.

項目Item型號PartUS$
藍牙主控晶片Bluetooth chipNordic nRF52832 · 1k2.67
藍牙模組（含天線）Bluetooth module (antenna)Ebyte E73 · 1004.59
加速度計AccelerometerST LIS2DW12 / Bosch BMI270 · 1k0.74 / 1.38
氣壓計BarometerBosch BMP580 / BMP3901.21 / 5.15
鋰電池LiPo cell250 mAh0.90–1.00
項圈元件小計（不含 UWB）Collar parts subtotal (no UWB)5.50–12
UWB（選配）UWB (optional)Qorvo DWM3000 / NXP SR04017–24 / ~10
家用鏡頭Home cameraTapo C210 / Furbo / RV1106 module~65 / ~54 / 40–60
房間接收器Room receiverESP32-C32.5–3
區域標籤Zone tagnRF52832 起，另加電池外殼nRF52832 + cell and housing2.67+
AI 配音（每句 50 字）AI voiceover (50-char line)—0.0002–0.002
雲端儲存與頻寬Cloud storage and bandwidth—未估TBD
10

對手各做一塊，第一年就收 119 至 370 美元Competitors each cover one piece and charge US$119 to 370 in year one

So what

飼主已習慣「硬體加訂閱」，月費落在 5 至 14 美元。多貓家庭是缺口：Catlog 按貓收費，兩隻貓每月約 28 美元；鏡頭類又分不出是哪隻貓。以「每戶」定價的整合方案有空間，定價仍待討論。Owners are used to hardware plus subscription, at US$5 to 14 a month. Multi-cat homes are the gap: Catlog charges per cat, about US$28 a month for two, and cameras cannot tell cats apart. An integrated offer priced per household has room; pricing is still open.

產品Product做什麼Covers硬體Hardware訂閱Subscription第一年（一隻貓）Year 1 (one cat)缺口Gap
Catlog (JP)感測項圈Sensor collar~$99每隻貓每月約 $14（亞馬遜標示 $8 起）~$14/cat/mo (Amazon lists from $8)$195–267沒有影像與區域；按貓收費No video or zones; charged per cat
Furbo 360° Cat寵物鏡頭Pet camera$54 (list $184)年繳平均每月 $6.99$6.99/mo avg, billed yearly~$138分不出個別貓咪Cannot tell cats apart
Petcube Cam 360寵物鏡頭Pet camera~$47每月 $5.99 起；進階 $11.99 不限鏡頭數From $5.99/mo; premium $11.99, unlimited cameras$119–191沒有訂閱只能看直播Live view only without a subscription
Petlibro Granary 2 VisionAI 餵食器AI feeder$190照護 $59.99／年；雲端錄影 $119.99／年Care $59.99/yr; video cloud $119.99/yr$190–370只看得到餵食器一個點Sees one spot: the feeder
Whisker Litter-Robot 4自動貓砂盆Litter box$699含一年 Whisker+ 的組合 $878$878 bundle incl. 1 yr Whisker+~$878只看得到貓砂盆一個點Sees one spot: the litter box
項圈攝影機Collar cameras (Whiskcam, Mr Petcam)第一人稱影片POV video未查Not checked無None—16 至 24 公克、無剪輯、無直播16 to 24 g, no editing, no live view
我們（第一階段）Ours (Phase 1)項圈＋鏡頭＋剪輯＋區域Collar + camera + clips + zones$46–77 元件parts待定To decide待定TBD元件成本未含電路板、外殼、組裝Parts exclude board, housing, assembly

價格為 2026 年 10 月美國官網或評測所列美元價；台灣售價與匯率待查。第一年＝硬體＋12 個月訂閱。Prices are US-dollar list or review prices as of October 2026; Taiwan pricing still to check. Year 1 = hardware + 12 months of subscription.

11

驗證計畫與 MVP：約 16 週的測試階段，先證明需求再證明技術Validation plan and MVP: a 16-week testing phase, demand first, then technology

So what

用現成鏡頭與人工剪片先測飼主是否「每天都想看」，同時平行驗證感測與區域。第 15 至 16 週以結果決定是否投入打樣，每一步都可能改變後面的方向。Use off-the-shelf cameras and hand-cut clips to test whether owners want this every day, while checking sensors and zones in parallel. In weeks 15 to 16 the results decide whether to fund prototypes; each step may change what follows.

週數為建議草案，待與 Rj 確認。Weeks are a draft proposal, to confirm with Rj.

步驟Step要回答的問題Question方法Method衡量什麼What we measure
1 需求驗證1 Demand哪 3 至 5 個時刻讓飼主每天想看、願意付費？Which 3 to 5 moments do owners want daily and pay for?現成寵物鏡頭；透過人脈找幾個多貓家庭；人工剪片、配音、做區域報告；請 KOL 發布測試Off-the-shelf pet cameras; a few multi-cat homes from our networks; clips, voiceovers and zone reports made by hand; KOLs post test clips每天可用片段數、報告開啟率、各時刻偏好、KOL 片段互動對比一般貓片、付費意願Usable clips per day, report open rate, preference by moment, KOL clip engagement versus ordinary cat videos, willingness to pay
2 鏡頭整合2 Camera候選鏡頭能否取得影像、依事件轉向與剪片？Can candidate cameras expose video, turn and clip on events?測試市售鏡頭的本地串流與開放介面；評估 RV1106 公版Test local streaming and open interfaces on retail cameras; evaluate an RV1106 board可行 / 不可行，與所需工程量Works or not, and engineering effort needed
3 感測器驗證3 Sensors跳躍、高度、互動偵測在真實家庭多準？How accurate are jump, height and interaction detection in real homes?開發板加加速度計與氣壓計給幾隻貓配戴，對照鏡頭畫面標註Dev boards with accelerometer and barometer on a few cats, labelled against camera footage準確率、誤報率、氣壓干擾、配戴接受度、耗電Accuracy, false alarm rate, pressure interference, wear tolerance, battery drain
4 區域驗證4 Zones視覺與藍牙哪條路徑誤報漏報較少？Which path, vision or Bluetooth, has fewer false and missed alerts?同一家中同時跑兩條路徑Run both paths side by side in the same home誤報、漏報、設定時間False alerts, missed alerts, setup time
5 代工詢價5 Quotes完整單件成本是多少？What is the full unit cost?帶著結果向深圳代工廠詢價Take results to Shenzhen factories電路板、外殼、組裝與總成本Board, housing, assembly and total cost
決策標準Decision criteria

開始前先和 Rj 訂好門檻（例如報告開啟率、付費意願、偵測準確率、單件成本上限），第 15 至 16 週依門檻決定：進入打樣、調整方向或停止。Set thresholds with Rj before starting (for example report open rate, willingness to pay, detection accuracy, a unit cost ceiling). In weeks 15 to 16 these decide: fund prototypes, change direction, or stop.

附錄APPENDIX

技術細節與資料來源Technical detail and sources

A. 技術可行性細節（Rj）A. Feasibility detail (Rj)

A1 動作與高度A1 Motion and height

  • 追逐（加速度計）：三軸資料以機器學習分類。文獻回顧：Galea 等人 2021 在 12 種行為約 98.9% 至 99.6%；Smit 等人 2023 在 15 種行為約 70% 以上。哪種模型最好仍是開放問題。LIS2DW12 約 0.74 美元，BMI270 約 1.38 美元。Chasing (accelerometer): classify 3-axis data with machine learning. Review: Galea et al. 2021 about 98.9 to 99.6% across 12 behaviours; Smit et al. 2023 above 70% across 15. The best model is still an open question. LIS2DW12 about US$0.74, BMI270 about US$1.38.
  • 追逐（鏡頭）：RV1106 等級晶片跑小型 YOLO 約每張 68 至 97 毫秒，約每秒 10 至 15 張；只在畫面內有效，速度受距離影響。Chasing (camera): small YOLO on an RV1106-class chip takes about 68 to 97 ms a frame, roughly 10 to 15 fps; works only in view, and speed depends on distance.
  • 跳躍：Sharon 等人 2020 實驗室中三種跳躍誤判率約 5.4%。氣壓計 BMP390 相對精度 ±0.03 hPa（約 ±25 公分），1 Hz 約 3.2 微安，約 5.15 美元；BMP580 約 1.21 美元。開門、冷氣、風壓干擾未查到資料，需實測。Jumping: Sharon et al. 2020, lab: about 5.4% misclassification across three jump types. BMP390 barometer ±0.03 hPa relative (about ±25 cm), about 3.2 µA at 1 Hz, about US$5.15; BMP580 about US$1.21. No data found on interference from doors, air conditioning or wind; needs testing.

A2 互動與個體辨識A2 Interaction and identity

  • 藍牙訊號強度：接觸追蹤研究均方根誤差約 3 公尺，身體遮擋影響大；另一研究 1 公尺內約 10 公分。掃描會增加耗電，需實測。Bluetooth signal strength: contact-tracing study about 3 m RMSE, heavily affected by body blocking; another study within 10 cm inside 1 m. Scanning adds power drain; needs measuring.
  • UWB：DW3000 空曠室內約 1 至 2 公分、遮擋時約 17 公分；DWM3000 約 17 至 24 美元，NXP SR040 約 10 美元；可能幾天就耗盡鈕扣電池，留給後續版本。UWB: DW3000 about 1 to 2 cm with clear sight, about 17 cm obstructed; DWM3000 about US$17 to 24, NXP SR040 about US$10; may drain a coin cell in days, so a later version.
  • 影像辨識個體：Whisker CatID 在裝置端做貓臉辨識（未公開準確率）；Petlibro Granary 2 Vision 可辨識最多 10 隻貓。自己做門檻不低，項圈識別碼是我們相對純鏡頭產品的優勢。Vision re-identification: Whisker CatID runs on-device face recognition (no published accuracy); Petlibro Granary 2 Vision recognises up to 10 cats. Building our own is not trivial; the collar ID is our edge over camera-only products.

A3 區域劃定與危險警報A3 Zones and danger alerts

  • 三種區域：偏好區（貓跳台、窗邊，也看出買的設備有沒有在用）、限制區（餐桌、特定房間，可只針對某隻貓）、極危險區（陽台、爐台、開啟的窗；停留超過設定秒數才觸發，飼主選只記錄、推播或警報聲）。Three zone types: preferred (cat tree, window perch; also shows whether gear is used), restricted (dining table, a room; can apply to one cat), danger (balcony, stove, open window; triggers after a set dwell time, owner chooses log, push or alarm).
  • A 純視覺：在鏡頭畫面上圈選。Tapo C210 已有自訂區域但只搭配一般動態偵測。限制：畫面外無法判斷；平面影像中「爐台上」與「爐台前」可能重疊（推論，需實測）。A Vision only: draw zones on the camera view. The Tapo C210 already offers zones but only with generic motion detection. Limits: nothing out of view; on a flat image "on the stove" and "in front of it" may overlap (inference, needs testing).
  • B1 區域標籤：MokoSmart M1 約 26×26×7 公釐、5.4 公克、CR2032 約 1.5 年；nRF52832 每秒廣播平均約 9.5 微安，200 毫安時約 2.4 年。B2 房間接收器：開源 Bermuda 以 ESP32-C3 每顆約 2.5 至 3 美元，三房約 20 美元以內；項圈只廣播，較省電。藍牙約房間級，牆壁約 ±2 公尺誤差。Catlog 基地台約涵蓋 60 平方公尺、四個項圈，但資料常缺漏。B1 Zone tags: MokoSmart M1 about 26×26×7 mm, 5.4 g, about 1.5 years on a CR2032; nRF52832 advertising once a second averages about 9.5 µA, about 2.4 years on 200 mAh. B2 Room receivers: the open-source Bermuda project uses ESP32-C3 boards at about US$2.50 to 3, under US$20 for a three-bedroom home; the collar only advertises, saving power. Bluetooth is about room-level, with walls causing about ±2 m error. Catlog's base covers about 60 m² and four collars, but data gaps are frequent.
  • C UWB：Qorvo DWM3001C 平面精度優於 15 公分，續航以月計；每房需三個以上基地台，留給高階版。對外說「區域級」，避免讓人以為是公分級。C UWB: Qorvo DWM3001C under 15 cm 2D precision, battery in months; three or more anchors per room, so a premium tier. Describe precision to owners as zone-level, not centimetre-level.

A4 拍攝、剪輯與配音A4 Filming, clipping and voiceover

  • 鏡頭：市售成品最快（Tapo C210 約 65 美元，可轉動但追蹤笨拙；小米 C300 依賴雲端、難整合）。代工公版自寫韌體較自主，可避開中國雲端（RV1106 約 0.5 TOPS，模組約 40 至 60 美元），但需韌體能力。Camera: retail is fastest (Tapo C210 about US$65, pans but tracking is clunky; Xiaomi C300 is cloud-dependent and hard to integrate). An ODM board with our own firmware gives control and avoids Chinese cloud (RV1106 about 0.5 TOPS, modules about US$40 to 60) but needs firmware skills.
  • 追蹤：內建動態追蹤會追任何會動的東西；構想中的「事件驅動轉向」是在設定區域時記住角度，事件觸發時轉過去（尚未查到現成產品）。Tracking: built-in motion tracking follows anything that moves; the proposed event-driven presets save an angle per zone and turn there on an event (no existing product found).
  • 擷取：本地循環錄影（Tapo C210 最大 512 GB 記憶卡），只上傳事件前後幾秒；藍牙約每秒更新，片段前後要留緩衝。全部上雲較簡單但成本未估。Extraction: local loop recording (Tapo C210 takes up to 512 GB cards), uploading a few seconds around each event; Bluetooth updates about once a second, so pad clips. Uploading everything is simpler but not yet costed.
  • 配音：ElevenLabs Flash 每千字約 0.04 美元；Google 標準語音每百萬字約 4 美元，Amazon Polly Neural 約 16 美元。中文語音質感與產生台詞的語言模型成本未查。Voiceover: ElevenLabs Flash about US$0.04 per 1,000 characters; Google standard voices about US$4 per million, Amazon Polly Neural about US$16. Chinese voice quality and the cost of writing lines with a language model not yet checked.
B. 資料來源B. Sources

市場與人口Market and population

競品與價格Competitors and pricing

通路Channels

技術與元件（Rj 原始來源）Technology and parts (Rj's original sources)