Pedestrian Detection Aided By Temporal Prior

In this paper, we developed an algorithm to construct temporal priors to improve the stability of pedestrian detections. The proposed temporal prior has no additional computation cost and could be applied to any image based pedestrian detector. We also present a FPGA implementation that can process VGA video at a frame rate of 30 similar to 40 frames per second in real time. The design is implemented on a Xilinx Zynq FPGA. It performs all steps of the algorithm in the FPGA fabric including: color space conversion, frame resealing, HOG feature extraction, and candidate evaluation in a sliding window. The design uses parallel pipelines to evaluate multiple scales per frame.