A hybrid fusion based frontal-lateral collaborative pedestrian detection and tracking

In the current context of autonomous or assisted driving, reliable pedestrian detection plays one of the key roles in performing situation awareness tasks. In this paper we propose a new method for pedestrian tracking based on hybrid sensor fusion. Pedestrian detection and tracking based on a single feature type cannot always achieve the desired performance, especially in low visibility conditions. To improve the pedestrian detection rate, in this paper we propose a method for high-level decision fusion based on the detection outputs from multiple cameras. Low-level processing entails the capture, calibration and registration steps. The high-level processing consists of pedestrian detection, tracking of the detected pedestrians, and fusing the tracking outputs into a common multi-camera view. By relying on parallel tracks we are able to detect objects and follow them more efficiently without losing track. Moreover, each target receives specific signature which serves as an identification of the individual target. When the target pedestrian from any of the side views enters the area of the front looking camera, the identity of its track is handed over to the corresponding front view tracker. The proposed strategy improves early detection of pedestrians, which is critical for driving safety.