Edge AI means running the analytics — the object detection, classification, or event logic — directly on the camera or a device right next to it, rather than shipping the video to a server or the cloud to be analysed. The camera's own processor looks at the frames, decides "that is a person", and emits a small result. The video may never leave the device unless something interesting happens.
The advantages are speed, bandwidth, and privacy. Deciding on the camera avoids the network round-trip, so an edge result can arrive in roughly 25–100 ms versus several hundred for a cloud round-trip; it lets the camera send mostly metadata instead of a continuous video stream, cutting upload by around 99%; and it keeps raw imagery local, which eases privacy and residency concerns. On-camera silicon (an NPU) commonly offers a few TOPS of compute, enough for person/vehicle detection and simple rules.
The pitfall is expecting edge AI to do everything. A camera's compute, memory, and power budget cap how big and how accurate a model can be, so the heaviest tasks — cross-camera re-identification, large vision-language reasoning, training — still belong on a server or in the cloud. Edge AI is best understood as the fast, cheap first filter; the deployment decision (what runs where) is the real design problem, and the model internals belong to the AI for Video Engineering section.
The handbook · 7 volumes
Fora Soft Video Engineering Handbook
Edge AI keeps analytics on or near the camera, so only events travel and bandwidth stays flat as cameras grow. Our seven-volume handbook covers the whole video stack, and for edge AI start with Volume 6: deployment topologies, model artifacts, detection and tracking. Volume 7 adds serving with TensorRT, real-time budgets and cost per stream-hour. Written by Nikolay Sapunov, CEO at Fora Soft.
Deciding which analytics should run as edge AI on cameras and which on servers?
Tell us your camera hardware and the detections you need, and we will split workloads between edge AI and servers with model choices and hardware sizing. 250+ video products shipped since 2005.
