Multi-target Detection And Tracking Of Shallow Marine Organisms Based On Deep Learning

Shallow marine ecosystems are an important source of food supply for the ocean and the Marine area with the most intensive human activity.With the proposal and implementation of the strategy of maritime power and blue granary,higher requirements are put forward for the shallow marine biological breeding principle and technology.In recent years,the development of artificial intelligence and big data technology has advanced by leaps and bounds.It is of practical and theoretical significance to realize the monitoring of the survival status of shallow marine organisms and the construction of a large database of shallow marine ecological breeding.In order to solve the related problems of multi-object detection and tracking of shallow marine organisms,considering the detection accuracy and tracking accuracy in practical application,this paper improves the You Only Look Once(YOLO)model and Simple Online Realtime Tracking with Deep Association Metric(Deep SORT)algorithm.The specific work is as follows:(1)For the shallow marine biological target detection based on computer vision in the Marine environment,the YOLO v5 shallow marine biological target detection algorithm integrated with the attention mechanism is designed.Considering the imperfection of the shallow marine biological species in the existing public dataset,collect submarine videos and convert them into video frames,to the video frame image enhancement processing,expanding samples,and increasing sample diversity of video frame images,using the Labelimg software,made of shallow marine biological dataset;Redesigning the a priori box by using the K-means clustering idea,adapt to the self-made dataset scale;Considering the characteristics of shallow marine biological environment,adding the attention mechanism to the YOLO v5 feature extraction network,reduce the redundant information,redesigning of the network structure,improved the detection accuracy of the shallow marine biological target in the complex environment;Build the target detection and simulation environment,design of the ablation experiments,the experimental results demonstrate the effectiveness of the improved algorithm for target detection in shallow organisms,the average detection accuracy was improved by 3.2%.(2)Aiming at the object tracking problem of shallow marine organisms in marine environment,a multi-object tracking algorithm based on improved Deep SORT is designed.The improved YOLO v5 algorithms replaces the Faster Region-Convolutional Neural Networks(Faster R-CNN)as the detector of the Deep SORT tracking algorithm,the cascade matching strategy is adopted to solve the problem that targets cannot be tracked for the long time;applying Soft Non-maximum suppression(Soft-NMS)to optimize the generation process of bounding box to improve the detection accuracy of high-density shallow marine biological targets;build the target tracking simulation environment and design the simulation experiment,and the experimental results show that the multi-target tracking algorithm of shallow marine organisms based on improved Deep SORT reduces the number of biological target tracking id in marine environment,and improves the accuracy of multi-target tracking in shallow marine organisms.