Edge vision's hard part is the control loop: a 12% dead zone beat model tuning

A developer writing on dev.to has published the architecture of Netra, a closed-loop pan-tilt camera in which an ESP32-CAM streams frames, a YOLOv8 model finds the person in them, and the detection drives two servos that turn the camera to keep that person centered, with every component coordinated over MQTT ^[1]. The useful part of the write-up is not the detector. According to the author, proportional response plus a dead zone did more for tracking quality than any model tuning did ^[2].

Start with the buffering, because that is where naive builds fall over first. The firmware exposes an MJPEG stream at 640x480 and roughly 15 FPS ^[3], which leaves about 67 ms of budget per frame ^[4]. The author credits PSRAM-backed dual buffering with keeping that stream smooth instead of stuttering every time a frame is being read ^[5]. That is the whole lesson in one line: a single buffer means the reader and the writer contend for the same memory, and the visible symptom is a hitch on every capture, not an error message.

Above the camera, a FastAPI server pulls frames, runs YOLOv8 Nano, computes where the target sits relative to the frame center, and publishes servo commands ^[6]. A second ESP32 consumes those commands and drives the pan-tilt servos ^[7]. A React dashboard carries the live feed, a manual joystick, mode switches, alerts and health monitoring ^[8]. MQTT carries the traffic between them, with WebSocket pushing updates into the dashboard, which the author values for decoupling: the detector does not need to know how many consumers are listening ^[9].

The control loop is where the design earns its keep. Move the servo whenever the target is not perfectly centered and the camera jitters constantly, chasing sub-pixel noise and never settling ^[10]. Two rules fix it. Servo speed scales with the distance of the target from frame center, so an edge detection gets a fast correction and a near-centered one gets a nudge ^[11]. And a 12% tolerance band around center produces no movement at all ^[12]. On a 640-pixel-wide frame, 12% is roughly 77 pixels of permitted wander ^[13], which is a large error to accept deliberately and exactly why the jitter stops. The mechanics are MG90S metal-gear servos with roughly 0 to 180 degrees of pan and 30 to 150 of tilt ^[14], so the tilt axis has two-thirds the travel of pan ^[15].

The rest is scope. Three modes: manual, auto, and patrol across predefined waypoints ^[16]. A layer the author calls BASE, the Behavioral Anomaly Signature Engine, does multi-factor threat scoring against configurable thresholds and builds time-weighted patrol heat-maps ^[17]. The system is designed for ESP-NOW mesh handoff across up to eight camera nodes ^[18]. The dashboard uses stream-isolated rendering so React re-rendering sibling UI never tears down the video element ^[19], plus a 4-second auto-reconnect ^[20].

What to watch is what the post does not publish. There is no end-to-end latency figure, no detection accuracy, and no false-alarm rate for BASE, which is described by intent rather than measurement ^[17]. Eight nodes is stated as a design target, not a tested configuration ^[18]. And the 12% dead band is tuned against roughly 15 FPS ^[3][12]; the same tolerance on a slower loop, or a faster-moving target, is the first thing that would need retuning. The architecture is the deliverable here. The benchmarks are still owed.