A vehicle detection system for autonomous vehicles

Abstract

Reaching the level of fully autonomous driving is one of the goals for research nowadays. It’s believed that autonomous vehicles will make the roads more safe by eliminating the human error from driving process. Also, it will increase the mobility of vehicles and help those who couldn’t drive by themselves. One main task of autonomous driving is vehicle detection, as the vehicle shares the road with other vehicles, the driver needs to maneuver those vehicles without any collision. In this thesis, different approaches used to tackle this challenging problem are explored. One of these approaches is to extract features from the image and use a machine learning technique to train a classifier to distinguish features of vehicles. The Histogram of Oriented Gradient (HOG) feature is a well-known feature that was used in object detection. A new HOG based feature is introduced in this thesis that was proven to increase the discriminatory power of HOG feature. The new HOG variant used the compass mask to compute the gradient of image patch. The conventional HOG and compass HOG were used to train SVM classifiers and their results were compared. The comparison also included the classification results of CNN based object detector (YOLO). The newly introduced feature is used in a vehicle detection system. An enhanced exhaustive search algorithm was used to scan the image frame for vehicle. The enhanced algorithm made use of the calibration data of cameras to reduce the search space and parallel computing to decrease the computation time.