Tracking and Motion Cues for Rear-View Pedestrian Detection

This paper describes a new system for detecting pedestrians using the rear-view camera only. The system is based on \"Accelerated Feature Synthesis\" (AFS) part-based object detection which enables the detection of pedestrians across a wide range of appearance changes: upright and non-upright pedestrians, partially occluded pedestrians and children. In this paper we enhance the AFS by introducing an integrated system taking into account temporal cues for reducing the system error rates. We introduce two new algorithmic components: Non-maximal suppression (NMS) tracking which uses visual tracking and detection history to enhance detection, and local-motion features which help identifying independently moving objects. In addition we use a collected application-specific training data and make its test part available as a new benchmark. Compared to the previously published results, our integrated system reduces the false alarm rate (at 90% detection rate) by a factor of 24 and shows a promising capability of detecting pedestrians and children in arbitrary poses.