Why the edge
Most computer vision work still happens in the cloud, where GPUs are cheap and power budgets do not matter. But for real products — a tracker in the field, a camera on a machine, a sensor on a vehicle — the cloud is too slow, too expensive, and sometimes simply unavailable.
Edge vision means running detection and tracking on the device itself. The wins are latency, privacy, and reliability: a decision in milliseconds, no footage leaving the hardware, and no dependency on a network link.
From pixels to detections
A practical pipeline is capture, preprocess, infer, post-process. Capture is the sensor; preprocess is resizing and normalisation; inference is the model; post-process turns raw outputs into boxes and classes.
The mistake we see most often is reaching for the largest model first. On embedded targets, a smaller quantised model that runs at the required frame rate usually beats a bigger one that cannot keep up. Measure throughput before you trust accuracy numbers.
Tracking over time
Detections are per-frame; products care about objects, not frames. Tracking joins detections across time with a simple association step — nearest-neighbour matching or a Kalman filter for smoother trajectories.
The key is graceful failure. When an object is occluded or the detector misses a frame, the tracker should coast on prediction rather than dropping the identity. A little temporal smoothing goes a long way in the field.
← Back to blog