Research On The Pedestrain Detection,tracking And Track Planning Algorithms For UAV

It is a development trend for UAV(Unmanned Aerial Vehicle)to realize intelligent and autonomous task execution.With the development of deep learning,there is a lot of room for development by combining computer vision technology with UAV platforms.Thersfore,This paper focuses on the plan planning and visual perception,mainly from three aspects: path planning,pedestrian detection and multi-target tracking.The main research contents of this paper are as follows:1.An improved sparse A* algorithm is proposed.Improve the classic A* algorithm from four aspects: dynamically measure the heuristic function,balance the speed and accuracy of the algorithm;add the maximum turning angle constraint,limit the search direction of the A* algorithm in the search process for the UAV flight degree of freedom;Use the bidirectional search strategy to search from the starting point and the target point at the same time to improve the overall speed of the algorithm;use the uniform B-spline curve to smooth the planned route and reduce the sharpness and unevenness at the inflection point of the path.The simulation environment was established in to verify the effectiveness of the improved sparse A* algorithm.2.An iterative pedestrian detection algorithm based on Faster RCNN is proposed.Pedestrian detection has problems such as crowded environment,dense pedestrians,and mutual occlusion.Therefore,this paper uses the iterative detection method based on the Faster RCNN algorithm,and sends the previous prediction result as a historical feature to the detection model again,so that the model can detect new pedestrians.The goal is to reduce the missed detection rate of the model.In addition,it is improved from three levels: sample level,feature level,and target level.Use the balanced sampling strategy to improve the selection probability of difficult negative samples;use the BFP module to average and refine the feature maps of different layers,and fuses high-level semantic information and low-level spatial information;use the balanced loss function to increase the gradient contribution of the inline value and limit the gradient contribution of the outline value,then balance the classification and regression loss function.Finally,the attention residual module based on the DCN is used to give full play to the advantages of DCN convolution in extracting features of different scales and irregular deformation,and effectively deal with pedestrian detection in complex environments.We experimentally verify the performance of the algorithm.The results shows that the proposed algorithm performs high accuracy and strong robustness.3.An online multi-target tracking algorithm with KLT optical flow is proposed.First,when the target is occluded or difficult to detect,it is difficult to match detection boxes and target boxes predicted by Kalman filtering correctly.Introduce KLT optical flow to make up for the insufficiency of Kalman filtering and more effectively use target motion information for prediction.In addition,Deep SORT use a simple appearance feature extraction model,which causes the target to fail to assign the correct ID identity when it reappears after disappearing for many frames.In this paper,the more mature appearance feature extraction models MGN and Mobile Net-V3 are used to replace the original models.We experimentally verify the performance of the algorithm.The results shows that the improved algorithm has higher tracking accuracy,and can assign the correct ID identity when the pedestrian reappears after occlusion,and has strong tracking performance.