Research On Moving Object Tracking And Velocity Measurement Based On Binocular Vision

Tracking and speed measurement based on binocular vision are widely applied in fields such as intelligent surveillance,autonomous driving,military reconnaissance,unmanned aerial vehicles,and more.When dealing with tracking and speed measurement problems in complex environments,binocular vision has become an ideal solution due to its advantages of low cost,small size,and strong environmental adaptability.The research in this article focuses on a binocular vision tracking and speed measurement system,which consists of two main parts: real-time target tracking using tracking algorithms and obtaining spatial position information of the target through stereo matching algorithms.The key challenge lies in ensuring real-time performance while accurately capturing the target and measuring its velocity.There are few publicly available algorithms that are both efficient and accurate.Therefore,striking a balance between real-time performance and accuracy is a crucial aspect of visual tracking and speed measurement algorithm research.To address this issue,this paper conducts research on target tracking and stereo matching,and ultimately establishes a visual tracking and speed measurement system based on the combination of Siam RPN++ and CGI-Stereo.For the target tracking task,this paper studies existing target tracking techniques and selects the Siam RPN++ algorithm as the baseline for the tracking algorithm.By introducing a pyramid-style feature fusion module,the shallow and deep features are combined,enhancing the network’s ability to represent the target.Simultaneously,the nonparametric attention mechanism Sim AM and spatial attention mechanism are introduced,enhancing the tracker’s target discrimination and anti-interference ability in complex environments.Compared to the baseline algorithm,the proposed algorithm improves the success rate by 2.3% and precision by 1% on the OTB2015 dataset,and success rate by0.9% and precision by 0.3% on the UAV123 dataset.Furthermore,the algorithm’s running speed is stable at 70 FPS.For the stereo matching task,this paper deeply analyzes the differences between traditional stereo matching algorithms and deep learning-based stereo matching algorithms,and compares their performance on the KITTI dataset.The running speed and matching results of deep learning-based stereo matching algorithms are then analyzed,with CGIStereo ultimately chosen to obtain efficient and accurate disparity maps.The tracking and speed measurement system studied in this paper uses the intersection over union(Io U)to mitigate the jitter problem of the target bounding box and introduces median filtering for post-processing of the disparity map,improving the accuracy of the disparity map and effectively reducing the matching point error value.In stationary ranging experiments,the maximum error within seven meters does not exceed 4%;in speed measurement experiments,the speed error is within 0.2 m/s.The running speed of the tracking and speed measurement system is stable at over 18 FPS.