Hardware Designs for Histogram of Oriented Gradients in Pedestrian Detection: A Survey

Abstract

Feature detection, narrowed down to pedestrian detection, is an imperative domain where automated applications such as Automatic Driver Support Systems, Robotics and similar image vision and machine vision technologies. Histogram of Oriented Gradients (HOG) is a robust, scalable and efficient feature extraction method that works on luminance gradients among neighboring pixels. The extracted feature is normalized and classified through support vector machines (SVM). Improvements in the design through approximate computations, parallelism and pipelining applied to SVM classification and histogram generation, Parallel implementation of entire HOG and exploration of possible applications of the algorithm. This paper cites the software improvements of HOG and hardware implementations targeted on FPGA for variations of HOG.