Anomaly Detection Of Fish Behaviors Based On Trajectory Extraction

China is a big fishery country and has always been the largest exporter of aquatic products.In the process of aquaculture,improper water environment will affect the growth of fish,and bring some abnormal behaviors.If these abnormal behaviors are not detected in time and no effective measures are taken,it will often lead to fish diseases and even large-scale deaths.Therefore,it is imperative to monitor fish behaviors.Based on the underwater fish video dataset,MHK,and the underwater fish trajectory dataset,Fish4 Knowledge,this thesis separately studies the technology of fish swimming trajectory extraction and the technology of abnormal trajectory detection,and finally integrates them into an online monitoring algorithm.For trajectory extraction,this thesis first trains Faster RCNN,an object detection model in deep learning,to detect underwater fish images.Data augmentation is used to enhance the model performance.A trajectory association method is then proposed,which defines the de-gree of relation between two bounding boxes and optimizes the results using temporal context information.For abnormal trajectory detection,this thesis first uses the GRU network(Gated Recurrent Unit)to design an autoencoder model for trajectory time series data,and transforms the trajectory data into fixed-dimensional feature through unsupervised learning.Then,two cri-teria based on reconstruction error and distance from cluster centers are proposed.Finally,this thesis combines the two parts of trajectory extraction and anomaly detection,and preliminarily completes an online monitoring algorithm.The results of the underwater fish test video show that the trajectory extraction method can correctly obtain the swimming trajectory of fish,reduce missed detection and false detection errors,and effectively improve the overall performance.The results of fish trajectory dataset show that the feature transformation method based on autoencoder is better than the method of artificial features,and the two anomaly detection criteria can effectively capture the abnormal trajectory.