Vehicle tracking using a generic multi-sensor fusion approach

This paper tackles the problem of improving the robustness of vehicle detection for advanced ACC and obstacle detection applications. Our approach is based on a multi-sensor data fusion for vehicle detection and tracking. Our architecture combines two sensors: a frontal camera and a 2D laser scanner. Improving robustness stems from two aspects. First, the vision-based detection by developing a multi-algorithm approach enhanced with a genetic AdaBoost-based algorithm for vehicle recognition is addressed. Then, the transferable belief model and evidence theory as a fusion framework to combine confidence levels delivered by the algorithms in order to improve the classification are used. The architecture of the system is very modular, generic and flexible: it could be used for other detection applications or using other sensors or algorithms providing the same outputs. The system was successfully implemented on a prototype vehicle and was evaluated under real conditions and over various multi-sensor databases and various test scenarios.