Dynamic Multi-vehicle Detection and Tracking from a Moving Platform

Recent work has successfully built the object classifier for object detection. Most approaches operate with a pre-defined class and require a model to be trained in advance. In this paper, we present a system with a novel approach for multi-vehicle detection and tracking by using a monocular camera on a moving platform. This approach requires no camera-intrinsic parameters or camera-motion parameters, which enable the system to be successfully implemented without prior training. In our approach, bottom-up segmentation is applied on the input images to get the superpixels. The scene is parsed into less segmented regions by merging similar superpixels. Then, the parsing results are utilized to estimate the road region and detect vehicles on the road by using the properties of superpixels. Finally, tracking is achieved and fed back to further guide vehicle detection in future frames. Experimental results show that the method demonstrates significant vehicle detecting and tracking performance without further restrictions and performs effectively in complex environments.