Moving Object Distance Estimation Method Based On Target Extraction With A Stereo Camera

As a hot topic of computer vision research,the stereo vision can not only identify and locate the targets,but also obtain the depth information of objects.Estimation of the depth information based on the stereo vision has wide of applications in the industrial fields such as robot autonomous navigation and obstacle avoidance,automatic driving,3D reconstruction,and virtual reality.With the increasing maturity of target detection and visual distance estimation technique and the increasingly complicate application scenarios,the distance estimation or non-deep target detection of static targets can not satisfy the requirements of applications.In this thesis,target tracking is combined with the stereo vision to achieve the real-time ranging of mobile targets.The detailed work is as follows:To combat the problem of unstable object during the real-time distance estimation of mobile object with the stereo vision,the region of interest(ROI)in each frame is tracked with the kernel correlation filter(KCF)tracking algorithm to reliably capture the continuous target regions quickly.To achieve the real-time distance measuring,a ROI extraction method is proposed in this thesis,which uses the target tracking results to intercept the same position and size in the synchronized frames(left and right)so as to reduce computation during the stereo matching.Experimental results show that the proposed method can increase the distance measuring speed to 0.142 seconds per frame with the stereo camera at a resolution of 320 * 240.To combat the problem of measuring accuracy due to the matching error during the real-time distance measurement with the stereo camera,the linear interpolation-based computation of matching cost is integrated with the cost aggregation based on the multi-directional one-dimensional dynamic programming to ensure the accuracy of the initial disparity.A CIELab color space-based disparity refinement updates the disparity according to the color similarity,which might reduce matching error so as to improve the measuring accuracy.Experimental results show that the measurement accuracy of the improved matching algorithm can achieve measuring accuracy above 95% in the range of 88cm-300 cm with the baseline of 151 mm.In addition,the impact of the baseline length on the distance estimation accuracy is studied by experiments,which provides a basis for the future work with binocular distance measurement.