· 5 min read · EdgeAI.computer Editorial
On-camera analytics or a separate edge box?
Cameras from Axis and Hanwha Vision now classify objects on the device. That changes where the inference workload should sit.
On-camera inference has moved well past motion detection. The Axis Q1656-DLE runs AXIS Object Analytics to classify humans, vehicles and unknown objects with deep learning, and fuses those detections with an on-board FMCW radar operating at 61.00–61.50 GHz, giving range, direction and velocity with detection up to 90 m in an area profile and up to 150 m for vehicles on a road profile. Hanwha Vision's XNB-9003 uses an on-board NPU for object, face and vehicle-type detection plus licence plate recognition on a 4K stream.
Where the camera can classify, the network carries events instead of video, and privacy exposure drops because imagery need not leave the device. Both vendors also publish bandwidth-reduction features — Zipstream on the Axis side, WiseStream III on the Hanwha side — which matter when a site's uplink is the real constraint.
A separate edge box still wins in three situations: when you need a model the camera platform cannot host, when you must correlate several cameras or other sensor data, and when you are retrofitting cameras that are already installed. That is the case a platform such as the ADLINK DLAP-411-Orin addresses, aggregating up to eight cameras over PoE and USB into one Jetson AGX Orin system.
In practice most sites end up with both: classification on the camera where the platform supports it, and a local system for aggregation, recording integration and the models that do not fit. Confirm which analytics applications a given camera model is certified to run before assuming the split.
Last updated .