Object-level fusion and confidence management in a multi-sensor pedestrian tracking system

This paper describes a multi-sensor fusion system dedicated to detect, recognize and track pedestrians. The fusion by tracking method is used to fuse asynchronous data provided by different sensors with complementary and supplementary fields of view. The confidence in detection and recognition is calculated based in geometric features and it is updated using the transferable belief model framework. The vehicle proprioceptive data are filtered by a separate Kalman filter and are used to estimate the relative and the absolute state of detected pedestrians. Results are shown with experimental data acquired in urban environment.