| Target identification and tracking technology is widely used in many fields.Traditional fixed monitoring equipments have a limited monitoring range and cannot track multiple targets for a long time.with the development of UAV technology,UAV has become a new type of monitoring equipment.For the target identification and tracking task from the perspective of the UAV,in addition to the traditional difficulties such as occlusion,it also faces challenges such as small target size and lack of computing resources.In this paper,the above problems are studied and a multi-target pedestrian identification and tracking algorithm suitable for drone platforms is designed.The main work of this paper is as follows:（1）In view of the problems of small target size and lack of computing resources from the perspective of drones,this paper is based on the SSD algorithm and improves the network structure,default box,loss function and post-processing.Firstly,mobile Netv2 is used to simplify the network structure,and a cascading feature fusion module is introduced to improve the detection effect of small objects.Secondly,the default box scale suitable for this scenario is analyzed by using K-means clustering algorithm.Then,focal Loss is used to calculate the classification loss,the improved DIOU Loss and the exclusion term are introduced to calculate the positioning loss.Finally,Soft-NMS is used for postprocessing.The experimental results show that the accuracy of the algorithm is improved by 7.47%,the speed is increased by 18 frames/s on Nividia Jetson TX2,and the detection effect of small targets and occluded targets is significantly improved.（2）Aiming at the problems of ID switches during target re-entry,ambiguity of target identity recognition by multiple drones,and redundant feature extraction process in existing recognition frameworks,this paper proposes a feature-sharing recognition framework and a re-identification method based on improved trihard loss function and data augmentation.In order to reduce unnecessary feature extraction operations,this paper proposes a identification framework for detecting and re-identifying shared networks,and improves Trihard Loss,while using data enhancement to expand existing samples.The Experimental results show that the improved loss function and data enhancement bring about an improvement in model performance,and the accuracy of the model on the Market1501 dataset reaches 72.7%,and the accuracy rates of Rank-1 reach 79.5%.Compared with the traditional two-stage target identification framework,the framework proposed in this article has increased the speed by 26%.（3）In order to reduce the impact of false detections and missed detections on the system,this paper designs a real-time target tracking algorithm based on the Deep SORT algorithm.Firstly,the motion model in the filter is replaced with an interactive multimodel to improve the tracker’s ability to predict the motion state of the target.Secondly,extracting apparent features using shared network structures in the identification framework.The experimental results show that the accuracy（MOTA）of the improved tracking algorithm is increased by 2.8%,the accuracy（MOTP）is increased by 0.7%,and the speed is accelerated by 37%,and can achieve a real-time speed of 26 frames/s on TX2.The target identification and tracking method proposed in this paper can run in real time on the UAV embedded platform,identify targets accurately and track targets continuously and steadily.The research results in this paper provide more ideas for the realization of UAV autonomy and intelligence,and have high practical value. |
