Deep Learning Based UAV Target Detection And Tracking

With the continuous development of technology,the use of drones is becoming more and more widespread.By carrying high-definition cameras on drones,ground station users are able to observe the images taken by drones in real time through terminals.Meanwhile,deep learning technology is widely used in daily life with its excellent performance capability.In response to the current algorithm’s insufficient ability to detect small targets in the images captured by UAVs due to the small size of targets and missing textures,etc.In this paper,a vision-based target recognition detection and tracking network structure is proposed to improve the detection and tracking performance of UAVs.The research is as follows.(1)Based on the lightweight target detection network Tiny Yolov3,the original K-Means clustering algorithm is improved by using the clustering number K=9 and using IOU instead of Euclidean distance to re-clustering the prior frame,taking into account the reality of this paper.Improve the Tiny Yolov3 detection network: add 3×3 and 1×1 convolutional layers,and also add a new upsampling layer Upsample 2 after the fifth convolutional layer,and perform Concat connection operation with the feature map output from the eighth convolutional layer to get a new output layer: 52×52.Add a new feature pyramid to improve the detection ability of small targets.(2)Using Kalman filter algorithm,a detection network incorporating Kalman filter tracking algorithm is proposed under the premise of tracking the target.The detection frame of the target detection network is optimally matched with the tracking frame of the Kalman filter algorithm using the Hungarian algorithm.The detection is corrected using the corrected results,which improves the detection speed and detection capability of the target detection network.(3)A comprehensive simulation environment based on robot operating system ROS,simulation software Gazebo and autopilot software PX4 is built to validate the proposed target detection and tracking algorithm for comparison tests.The test results show that the improved Tiny Yolov3 network structure model improves the average detection accuracy m AP by 6.5%compared with the original Tiny Yolov3 network model.After incorporating the Kalman filter tracking algorithm,the average detection speed FPS of the detection network incorporating the Kalman filter tracking algorithm is improved by 34.2% and the detection accuracy m AP is improved by 8.6% compared with the improved Tiny Yolov3 detection network.The test results show that the detection network incorporating Kalman filter tracking algorithm can improve the network detection performance and has certain engineering application value.