Joint Human Detection From Static and Mobile Cameras

Abstract

Efficient pedestrian detection is a key aspect of many intelligent vehicles. In this context, vision-based detection has increased in popularity. Algorithms proposed often consider that the camera is mobile (on board a vehicle) or static (mounted on infrastructure). In contrast, we consider a pedestrian detection approach that uses information from mobile and static cameras jointly. Assuming that the vehicle (on which the mobile camera is mounted) contains some sort of localization capability, combining information from the mobile camera with the static camera yields significantly improved detection rates. These sources are fairly independent, with substantially different illumination and view-angle perspectives, bringing more statistical diversity than a multicamera network observing an area of interest, for example. The proposed method finds applicability in industrial environments, where industrial vehicle localization is becoming increasingly popular. We implemented and tested the system on an automated industrial vehicle, considering both manned and autonomous operations. We present a thorough discussion on practical issues (resolution, lighting, subject pose, etc.) related to human detection in the scenario considered. Experiments illustrate the improved results of the joint detection compared with traditional independent static and mobile detection approaches.