Implementation of Human Detection on Hardware

Abstract

This thesis presents a methodology to design an embedded vision system to detect upright humans in images and video feed from camera. Our goal is to develop an algorithm that can extract high dimensional feature vectors from encoded image regions that can substantiate object/non-object decisions. A simple learning framework is used to differentiate between an object and non-object in an image region using support vector machine as the classifier. A combination of both bottom-up and bottom-down approaches are used to design the system so that the benefits of both approaches be utilized to elucidate the algorithm. A detailed study of existing state of the art object detection algorithms is carried out, they perform well on general purpose computers with large memory units and adequate computational power but they fail to perform when it comes to implementation on single board computing devices with limited computational power and memory resources. The algorithm is based on HOG (Histogram of Oriented Gradients) features presented by N. Dalal and B. Triggs which is optimized to use efficiently in embedded hardware so that the results can be achieved in real time. A dense overlaid grid is formed over the image upon which histogram of oriented gradients is calculated keeping the resolution fixed, resulting in a high dimensional feature vector. The HOG descriptors are robust to significant variations in illumination, colour and minute variations in image contour locations and directions. ACF (Aggregated Channel Features), HOG features from literature and our implementation of HOG are evaluated on INRIA Person Dataset on a Intel Core i7 processor with 8GB RAM and on Raspberry pi B+. Almost all performed equally on dataset, but they have their own failures, like ACF performs well on desktop machine and is relatively faster but on single board computer it is slower than HOG which means that ACF is computationally more expensive than HOG. The algorithm developed is first implemented on MATLAB, then python implementation is completed on desktop machine which is then optimized and implemented