Detecting parallel moving vehicles with monocular omnidirectional side cameras

In this paper, we present a strategy for the detection and tracking of dynamic objects exploiting monocular omnidirectional side cameras. The main novelty of the approach is the use of solely motion based (optical flow) extracted image features from omnidirectional side cameras to continuously track parallel moving vehicles using a novel clustering algorithm. Firstly, optical flow features are extracted from side camera images. Secondly, these extracted features are identified as belonging to dynamic obstacles via positive-depth, positive-height, and epipolar constraint. A new method for constraint evaluation on omnidirectional cameras is presented, incorporating uncertainties of ego motion measurements. The features are clustered based on spatial closeness and optical flow similarity. Results of experiments, with real sensor data from a test vehicle, are presented.