Visual Object Tracking on the Inverse Perspective Map for Autonomous Vehicles

Visual object detection and tracking is a crucial element of Autonomous vehicles. Vision-based sensors are used to detect road entities such as pedestrians and vehicles, and track them in real-time. Various filters such as Kalman Filters and Particle Filters are used in tracking. However, the tracking of these objects is predominantly done on the image plane. The output from the camera is a perspective image of the environment, where the same number of pixels along different directions of the image may not correspond to the same distance in real-world units. Thus, any motion model assigned to the entity would not capture the exact dynamics of the object as the object moves in the 3D world but is being tracked on a perspective view. In this paper, a system is proposed to track the objects on the Inverse Perspective Map, which is a birds eye view of the environment obtained through a homography transformation of the perspective image. Experiments are conducted to show the working of the technique and the results are presented.