FPGA-Based Selected Object Tracking Using LBP, HOG and Motion Detection

This paper describes the hardware architecture for selected object tracking on an embedded system. The LBP and HOG feature extraction algorithm is combined with motion detection to compute and compare the features vectors with captured once only when the target moves. LBP8,1, LBP16,2, and HOG(8,1), HOG(16,2) are used to create the feature vector. The unit that makes a final decision on tracker update is based on searching of the least SSD of features' histogram. The implemented motion detection algorithm was able to find and mark eight moving objects simultaneously. The previously computed locations update all trackers' locations in every next frame. The experimental investigation showed that implemented tracker, based on HOG features is robust to luminescence variation and partial occlusion. In addition, the LBP based tracker is robust to the rotation. The proposed architecture is implemented on Xilinx Virtex 4 FPGA using VHDL and is able to work in real-time on 60 fps and 640x480 video resolution.