Research On Animal Target Recognition And Tracking Method Based On Deep Learning

Animal object detection and tracking have been a hot research topic in the field of computer vision for a long time.With the continuous development of computer hardware and deep learning technology,research on animal object detection and tracking is receiving increasing attention.The challenge of animal object detection and tracking lies in the fact that the appearance,shape,color,and movement of animals can vary greatly,and there are also significant differences between different species.In addition,due to the complex and changing ecological environment,factors such as lighting,weather,and occlusion can affect image quality and animal recognition,making animal object detection and tracking even more challenging.In recent years,with the development of deep learning technology,especially the continuous improvement of object detection and tracking algorithms,the performance of animal object detection and tracking has been greatly improved.The YOLOv5 object detection algorithm and the Deep SORT multiobject tracking algorithm based on object detection have achieved excellent results in pedestrian tracking,but research on animal object tracking is relatively limited.This thesis conducts in-depth research on the problems of YOLOv5 and Deep SORT algorithms in animal object detection and tracking,and optimizes and improves them based on these problems.The main research contents of this thesis are as follows:(1)To address the problem of insensitivity to small targets when using the Intersection over Union(Io U)metric to measure detection accuracy,a target detection evaluation method based on Wasserstein distance is introduced,and the target positioning loss in the YOLOv5 training process is optimized and improved based on this method.In addition,the coordinate attention mechanism is fused with the feature pyramid structure of YOLOv5 to enhance the network’s feature extraction ability for animal targets,thereby improving detection accuracy.Experiments on the dataset constructed in this thesis show that the mean average precision(m AP)metric has increased by 2.8 percentage points,which can effectively improve the detection effect.(2)To address the problem of trajectory mismatch in the animal object tracking process using the Deep SORT algorithm,Shuffle Net is used as the appearance feature extraction network,and the ECA attention mechanism is added to optimize the network’s feature extraction ability.In terms of trajectory matching,the DIo U(Distance-Io U)that comprehensively considers the overlap degree and distance between target boxes is used to improve the matching process.Experiments on a self-made dataset show that the Multiple Object Tracking Accuracy(MOTA)metric has increased by 3.7%,and the multiple object tracking precision(MOTP)metric has increased by 2.6%.The results show that the improvement proposed in this thesis can effectively solve problems such as trajectory disappearance and target identity ID switching in the animal object tracking process.(3)Based on some possible scenarios in real life,such as ecological conservation and agricultural breeding,an animal detection and tracking system is designed and implemented,and the animal object detection and tracking algorithm proposed in this thesis is applied in practical applications.The system can detect and track animal targets in video or live streams,visualize the tracking results,and provide users with convenient historical record management functions.