Moving Object Detection with Photometric Monocular SLAM on a Moving Ego-Platform

This paper describes a novel method for detecting moving objects using wide field of view monocular cameras and Direct Simultaneous Localization And Mapping (D-SLAM) on a moving platform. The main idea is to use (1) the quality of photometric patch matching and (2) deviation from the epipolar 3D constraint matching to separate static from dynamic objects. The result of the SLAM-based detection is a set of points, which are clustered together as boxes and tracked using an Extended Kalman filter. In order to relate the track to car coordinates, a ground detection algorithm used to extend the tracking box to the ground level for a proper distance estimation. No training data is required for the process and the algorithm has been tested on several different cameras from different manufacturers. A demonstration system running in different automotive platforms has been developed and evaluated, where detection accuracy is significantly improved compared to previous arts as demonstrated through precision-recall graphs. The algorithm is runnable at full frame rate at 30Hz on a Raspberry Pi ARM or an automotive TI TDA4 processor, suitable to be deployed in real automotive applications.