Research On Underwater Fish Tracking Based On Deep Learning

Fish is an important indicator of water quality.The change of fish behavior can timely and effectively reflect the change of water quality.Fish tracking is the basis for analyzing fish behavior,which can be used to analyze fish swimming speed and swimming trajectory.The data of fish swimming speed and acceleration can reflect the movement vitality of fish in a specific environment,and its swimming trajectory can analyze whether fish adapt to the environment to a certain extent.Therefore,it is of great significance to study the tracking method of underwater fish.However,there are many problems in the current fish tracking algorithm.Firstly,most of the algorithms in the field of fish multi-target tracking are non-real-time algorithms,which can not give consideration to speed and accuracy.Secondly,there is less research on fish tracking algorithm with small samples.Therefore,in order to solve the problem of insufficient real-time performance of current fish tracking algorithms and the problem of fish tracking in the case of small samples,this dissertation proposes two solutions:(1)This dissertation proposes a real-time fish tracking scheme based on the TBD(Track by Detection)framework,that is,the fish detector is used to detect the fish in each frame image,and then the tracking algorithm is used to track the detected fish across frames.In the detection stage,Yolov4-tiny RFB detector is proposed for fish detection.Yolov4-tiny RFB detector is based on Yolov4-tiny network,and RFB(receptive field block)module is added to the network,so that the detector can integrate the features of different scales,take into account the detection of targets of different sizes,and improve the detection performance.In the matching stage,according to the characteristics of fish non-linear motion,the Sort(Simple Online and Realtime Tracking)algorithm is optimized to improve the tracking accuracy.In this dissertation,the optimized algorithm is named Simple-Sort algorithm.Finally,a red snapper fish detection and tracking dataset in a real breeding environment with target box annotations is made for experimental verification.(2)A few samples underwater fish tracking method based on semi-supervised and attention mechanism is proposed,and fish detection and tracking are carried out by using labeled and unlabeled data.Inspired by the knowledge distillation method,this dissertation designs a novel semi-supervised Self-training method to train fish detection network.In this dissertation,pseudo-labels are used as the medium of knowledge transfer between teacher network and student network.Pseudo-labels generated by teacher network are used to guide the training of student network and complete knowledge transfer.In this dissertation,Yolov4 model is used as teacher network and Yolov4-tiny CBAM model is designed as student network.Yolov4-tiny CBAM is based on Yolov4-tiny network and adds CBAM attention mechanism to assign more weight to more important information,and suppress invalid feature information to make it learn more valuable information in limited samples.The Yolov4-tiny CBAM model based on semi-supervised training is combined with Simple-Sort tracker to realize the detection and tracking of underwater fish under the condition of few samples.The experimental results show that the two solutions proposed in this dissertation can meet the real-time requirement of 25 FPS for surveillance video.In the first scheme,the improvement of the Yolov4-tiny detector and the Sort tracker can effectively improve the detection and tracking accuracy,taking into account the accuracy and speed,but it needs to spend a lot of cost to make fish datasets.The second scheme,with the help of semi-supervised training and attention mechanism,can achieve near full-supervised performance with only a small number of samples.In conclusion,the scheme proposed in this dissertation can realize real-time and accurate tracking of fish in aquaculture environment.