FPGA implementation of HOG based multi-scale pedestrian detection

Pedestrian detection is needed for many vison applications including surveillance, Advanced Driver Assistance Systems (ADAS), Intelligent Transport System (ITS), drone, robotics, etc. There are different sizes of pedestrian or human in an image due to the different distances from the camera and different object's height. To detect all of the objects with different sizes, a multi-scale detector is needed. In this study, a multi-scale pedestrian detection is designed based on histogram of oriented gradients (HOG) and implement the method on a field programmable gate array (FPGA). The processing includes three stages: the input color image is converted to a gray one and then down sampled the gray image by 2 and by 4, respectively. different window sizes are used to extract the features of HOG from three size gray images. Final, linear support vector machine (SVM) is used to classify the extracted features for different window sizes. The experimental results show that the system costs 94,374 logic elements, which is about 82% of total logic elements s of Terasic DE2-115 development board. The system detection accuracy is about 97% on average and the processing speed can achieve 60 fps for 640×480 resolution.