Research And Implementation Of Multi Object Automatic Detection,Tracking And Recognition Algorithm

While real-time multi-object tracking(MOT)on UAV platforms is one of the most practical research directions in UAV research,it faces many challenges.At present,mainstream multi-object tracking algorithms all use the tracking by detection(TBD)paradigm,which first uses the object detection network to find out all the objects that need tracking and then track.Since the height at which the UAV acquires images is often far from the ground,the objects to be tracked in the images are usually small and dense,which means it is necessary for the object detection algorithm to have high accuracy.Also,real-time MOT algorithm calls for real-time performance,so its operating speed should be fast.And due to the limitation of the platform,it is hoped that the computational complexity of the algorithm can be as small as possible.This thesis proposed an improved YOLOv4 object detection algorithm and then combines it with a MOT algorithm,Deep SORT,to meet the above requirements.The main contributions of this thesis are:(1)An improved object detection algorithm is proposed.Based on the state-of-theart object detection algorithm YOLOv4,this thesis has improved its backbone,neck,prediction head and data augmentation methods.Specifically,optimization measures adopted are: adding a Focus structure to the input of the v4 backbone to retain more detailed features of the input image;replacing the Residual structure in the backbone with a Dense structure to improve network performance while reducing network parameters;the CSP structure is added to the neck to make the feature fusion better;during the test,the image adaptive scaling is used to speed up the inference speed while testing.The improved algorithm is called the CSPDense YOLO algorithm.An experiment was designed to compare the performance of the algorithm with the YOLOv4 algorithm on the Vis Drone2019 object detection dataset.The experimental results show that the proposed CSPDense YOLO algorithm on the UAV object detection data set has a 6.1%increase in m AP compared to YOLOv4,and FPS has been increased by 6.2%.(2)An improved multi-object tracking algorithm is proposed.Based on the Deep SORT algorithm,the feature extraction network of it is replaced with a faster and more accurate CSPDense YOLO object detection network.Further,the detection head of the network is modified for the needs of multi-object tracking tasks,so that it outputs the position,classification and appearance features of objects.The output object position information is used for moving object modeling,so as to calculate the motion feature similarity between the detected object and the tracked object;the output appearance feature,that is,embedding information is used to calculate the appearance similarity;the two similarities are fused to obtain the feature fusion similarity.Then use the cascade matching algorithm to alleviate the ID switching phenomenon while tracking.An experiment was designed to compare the performance of the improved algorithm on the Vis Drone219 multi-object tracking data set.The experimental results show that compared with Deep SORT’s MOTA,the proposed algorithm is improved by 10.1%,MOTP is improved by 10.3%,the number of ID switches is dropped by 24.5%,and FPS increased by 52.1%.(3)Integrated the proposed algorithm on the NVIDIA Jestson TX2 platform,and use Tensor RT to optimize and accelerate the proposed CSPDense YOLO algorithm.Then the algorithm was tested on the platform,and the FPS of the algorithm was as high as 24.2frames per second;the actual measurement was carried out on the UAV platform,and the results proved that the performance of the algorithm on the UAV platform was sufficient to meet the practical needs.