Real-time Animal Detection System for Intelligent Vehicles

Animal and Vehicle Collisions (AVCs) have been a growing concern in North America since the abundant wildlife resources and increases of automobiles. Such problems cause hundreds of people deaths, thousands of human injuries, billions of dollars in property damage and countless of wildlife deaths every year. To address these challenges, smart cars have to be equipped with Advanced Driver Assistance Systems (ADAS) able to detect dangerous animals (e.g., moose, elk and cow), which cross the road, and warn the driver about the imminent accident. In this thesis, we explore the performance of different image features and classification algorithms in animal detection application, and design a real-time animal detection system following three criteria: detection accuracy, detection time and system energy consumption. In order to pursue high detection rate but low time and energy consumption, a double-stage detection system is proposed. In the first stage, we use the LBP adopting AdaBoost algorithm which provides the next stage by a set of region of interests containing target animals and other false positive targets. Afterward, the second stage rejects the false positive ROIs by two HOG-SVM based sub-classifiers. To build and evaluate the animal detector, we create our own database, which will be updated by adding new samples. Through an extensive set of evaluations, we note that the double-stage system is able to detect about 85% of target animals.