Vision based long range object detection and tracking for unmanned surface vehicle

Abstract

A real time vision based long range object detection and tracking algorithm for unmanned surface vehicles (USV) is proposed in this paper. HD image (2736 × 2192) is utilised in this work to obtain high accuracy for the object distance estimation. With handling such high resolution images for real time performance, we propose a coarse to fine approach, which firstly estimates the sea surface plane and locations of objects coarsely on lower resolution images corresponding to the HD images, then the detected coarse locations or regions of interest (ROI) are projected to the original HD image, finally stereo matching is preformed in the original image only on these extracted ROI, which renders more accurate 3D information for localizing the objects on the open sea. In the tracking, we propose to combine the target tracking based on 2D image with the constrained template matching to compute the depth, which demonstrates a more robust and accurate performance. Experimental results with our own dataset verify the high efficiency of our proposed method.