Research On Vision-based Remote Target Detection And Tracking Technology

With the gradual maturity of the visual field and the continuous improvement of computer computing capabilities,urban surveillance,smart devices,More and more scenes such as unmanned driving have a demand for target detection and target tracking,so many challenges are also ushered in.Most target detection and tracking systems are deployed on local local area networks,and it is difficult to obtain remote video streams and complete detection and tracking tasks in real time.In the target detection link,there are greater challenges for target detection in scenes such as small targets,occlusions,and dense targets;similarly,in the target tracking link,when the target is occluded,,motion blur,etc.,target loss is prone to occur Or the positioning is not accurate,and it is difficult to achieve long-term tracking.In this paper,a set of remote target detection and tracking system is designed for the purpose of realizing long-term tracking of targets and improving the robustness and real-time performance of target detection and target tracking.The system adopts C/S architecture and consists of remote communication module,target detection and tracking module,and mobile robot Navigator Q2 control module.First of all,in order to meet the video stream frame rate and image quality requirements,this article uses remote networking to realize remote communication,virtualizing the different local area networks where the industrial computer and the remote PC are located into a new local area network,and based on the TCP/IP network protocol,Socket network programming is used to realize the transmission of video stream and control commands.Secondly,the target detection selects pedestrians as the detection target,and the YOLO v3 algorithm is used as the algorithm prototype.Through the refinement of grid division,the situation that the center positions of some targets are in the same grid is improved to reduce the occurrence of missed detection;based on the more typical width and height of pedestrians The prediction of candidate frames is improved,and the Euclidean distance is replaced by Io U to obtain a more accurate clustering center and pedestrian aspect ratio;the receptive field enhancement module is added to improve the network structure to increase the receptive field of var Io Us scales,so that the detection algorithm is suitable for small targets and The occluded target can be positioned more accurately.For the target tracking algorithm,select the KCF algorithm as the prototype,add a one-dimensional scale filter and select the scale that obtains the largest response as the scale of the tracking frame,use FHOG,CN dual-core feature fusion,increase the feature extraction capability;no longer use interpolation as an update Strategy,based on historical PSR,enables the model and parameters to be updated adaptively,and enhances robustness to occlusion situations.The target detection and tracking module combines the detection and tracking algorithms by adding a re-detection module to enhance the self-correction ability in the tracking process.Then,a manual control algorithm is designed based on the control protocol of the Navigator Q2 robot platform;a set of follow-up control algorithm is designed based on the scale and position information of the target tracking result frame,including the target search strategy after the target is lost,which greatly improves the system's performance Long-term tracking capability.Finally,through experiments on each module and the system as a whole,it is proved that the algorithm improvement and the effectiveness of the system are feasible,which have practical significance.