Research On Multi-Object Tracking Algorithm Based On Deep Learning

Multi-object tracking(MOT)is one of the important research directions in computer vision.In recent years,with the development of deep learning,MOT has also made significant progress.However,in the drone scenario,MOT still faces many challenges,such as low detection accuracy due to small objects,low identity consistency due to severe occlusion,limited computing resources leading to inability to flexibly deploy deep learning models,etc.In this thesis,we propose the following improvements to address these issues:1.To address the problem of low accuracy in detecting small object in drone images,this thesis proposes an object detection backbone network based on a large convolutional kernel architecture,built on top of YOLOv7.By increasing the effective receptive field of the model,the proposed architecture enhances the feature extraction capability,thereby improving the detection accuracy.Moreover,to reduce the additional parameters brought by the large convolutional kernel,this thesis uses techniques such as Struct Re-parameterization and Depthwise Convolution to reduce the model parameters.2.To further improve the model’s real-time performance,this thesis analyzes the sturct of different Feature Pyramid Network(FPN),and inspired by the Bi FPN structure,proposes an improved FPN struct on top of YOLOv7.The proposed feature fusion network enhances the information interaction between different-scale feature maps through Cross-Scale Connection.It can fuse more features with fewer parameters than other feature fusion networks.3.To address the problem of weak data association ability of OC-SORT,this thesis proposes an improved OC-SORT method.On one hand,it uses a real-time object detector based on large convolutional kernel architecture to improve the detection accuracy of small objects;on the other hand,it proposes a method that uses Re-identification and Kalman Filter cascade to enhance data association.Through experimental verification,this thesis’ s method can better improve identity consistency in dense scenarios.4.To address the problem of limited computing resources in drones,this thesis proposes a cloudnative deep learning container deployment platform,which builds a lightweight Kubernetes platform.On one hand,it converts models to ONNX format and encapsulates them as Docker containers to solve the problem of difficulty in managing deep learning frameworks,and also reduces the container image file size;on the other hand,it uses distributed container image repository and Bit Swap protocol to solve the problem of fierce competition for bandwidth at traditional centralized image repository exit,greatly improving the efficiency of edge-side deep learning container deployment represented by drone container deployment.