Couch Presence Sensor

① The problem.

I want the TV to pause when I get off the couch and resume when I sit back down — no remote, no phone. The common sensors all miss the obvious case: PIR and most radar see motion, so a person sitting still to read scans as an empty room and the TV pauses on you. And my couch sits in the middle of the room — an island — so a room-level sensor can't tell "on the couch" from "walking past it."

② Approach.

First attempt: put the detection on a camera aimed tightly at the couch. A Seeed XIAO ESP32-S3 Sense running on-device person detection — esp-dl with Espressif's pedestrian-detect model — so "person in frame" means occupied, the still-person case that motion sensors fail at. Aiming tight at the couch dodges the island problem. The board feeds raw RGB565 frames to the model, debounces (one second to confirm presence, eight to confirm absence so a dropped frame doesn't pause you mid-show), and publishes ON/OFF to an MQTT topic. Home Assistant turns that into a binary sensor, and two automations pause and resume the LG webOS TV. Everything except the sensor itself — broker, binary sensor, automations, OTA, debug stream — is sensor-agnostic by design.

③ What's in the box.

• Sensor (v1) — XIAO ESP32-S3 Sense, OV2640, running esp-dl plus pedestrian-detect fully on-device. No frames leave the board; it publishes only ON/OFF.
• Messaging — retained presence plus an availability last-will to the self-hosted MQTT broker.
• Home side — a Home Assistant MQTT binary sensor and two automations driving an LG C5 OLED over webOS. The triggers use a bare to: on/off plus a playing/paused condition, so a sensor reboot blipping to unavailable doesn't wedge the resume.
• Debug — an on-demand stream with the detection box drawn, a single-shot snapshot endpoint for remote framing, and HTTP OTA over dual partitions. Invaluable for aiming a headless camera from across the house.

④ What broke.

The vision model was the wrong sensing organ — and finding that out took two layers of debugging. First, it only works on raw RGB565 fed straight to the model: every JPEG-decode path I tried (three different decoders) returned zero detections, and aggressive sensor gain washed the frame flat green and starved it. Sane exposure bought me one good pause/resume cycle.

But the deeper problem is the model itself. It's trained on upright pedestrians, scores a seated person at a marginal 0.74–0.81 against its 0.70 threshold, and largely fails on the postures a couch is for — slouched, leaning, lying down, hunched over a bowl. A dimming evening room tips it to zero. No threshold tuning fixes "this model has never seen a person lounging." So I'm pivoting the sensor to an LD2450 mmWave radar, which sees a body in any posture, in the dark, holding perfectly still. Because the plumbing was built sensor-agnostic, only the firmware's detection section changes — the broker, Home Assistant, and the automations are zero-touch.

⑤ Where it's going.

The LD2450 runs on a spare ESP32 WROOM under ESPHome's native component — radar needs only UART and WiFi, so it frees the camera-class S3 entirely (which got repurposed into a rep counter). It publishes to the same MQTT topic, so Home Assistant never learns the organ transplant happened. The ready-to-flash plan and full ESPHome config are written and waiting on the board. Later: zone-based X/Y so "on the couch" is a literal box in the radar's field, and maybe gesture control once presence is rock-solid.