Research On Fish Detection And Tracking Algorithms In Natural Waters

Vision-based fish detection and tracking in natural waters is of great significance for improving the level of informatization,automation,and intelligence in the process of aquaculture,and is of excellent application value for fish behavior analysis,species observation,and water quality monitoring.To solve the problem of the fish swimming posture changes,uneven illumination,complex background,and lack of fish data set,in this paper,we mainly research:Firstly,Research on a fish detection algorithm based on improved GBVS.GBVS is a selective attention mechanism by simulating humans,first of all,according to the fish characteristics extract phase consistency feature,color name features and illuminance feature;Then,the Markov chain is used to calculate the saliency value of each feature map separately,and the generated activation map is normalized,and merged into a final saliency map by linear superposition;finally,threshold segmentation and morphological operation are used to realize fish localization and segmentation.Secondly,Research on the improved Minimum Output Sum of Squared Error tracking algorithm are used to real-time fish tracking in natural waters.For the translation estimation,we extract gray feature and color feature of target,respectively.And then train the filter function to get their response value,the two kinds of response map are adapted to the adaptive fusion.For the scale estimation,the scale of multi-scale image space is estimated by the color and shape feature of fish.Avoid scale expansion or contraction,the consistency check is used.Thirdly,research on the long-term fish tracking method based on the framework of tracking learning detection in natural waters.In the tracking module,using scale adaptive mean shift replace the TLD tracking module;In the detection module,when the fish tracking is successful,the shape features of fish are added near the target in the previous frame to realize detection.Fourthly,In the absence of underwater fish tracking data set,video photographed in natural waters was collected from video website and manually annotated.Using the data set,the improved GBVS detection algorithm result shows that the method can realize fish detection.According to the experimental results of the improved MOSSE tracking algorithm,the method can realized fish track in real time.Experimental results of the improved TLD tracking algorithm show that this method can achieve long time fish tracking.