AI-based Multi-target Morphology And Trajectory Tracking Technology And Its Application In Water Conservancy And Ecological Monitoring

In this study,the AI-based multi-target shape trajectory tracking technology solves the problem that data such as species,quantity,size,swimming speed and direction of fish in the fishway are difficult to obtain comprehensively and continuously at same time.It has been noted that building dams can affect fish migration,which in turn affects communication between fish populations.In line with the concept of protecting the natural ecological environment and promoting the improvement of fish passages,people often collect fish data to evaluate the impact of engineering construction on the effect of fish migration.Therefore,a full-cycle,multi-faceted and integrated collection of fish data is required.However,the implementation of this statistical research process is difficult and cumbersome.It is difficult for researchers to obtain all-round fish passage information in the fishway,and artificial intelligence technology provides new solutions for obtaining these data.In this thesis,the method based on real-time analysis of video recording is used to collect and record fish data in the whole days.In terms of fish individual identification,fish tracking and data statistics,this study uses the improved YOLOv3 algorithm based on deep learning combined with the Deep-Sort method to identify and record fish information and achieves the effect of simultaneously recording multiple aspects of fish data information.The improved network structure in this study is more sensitive to small target objects,and the detection speed is faster,and the model recognition accuracy rate reaches 94.6%;the proposed fast dataset expansion method can shorten the model training time and speed up the project deployment.The full-time and multi-type fish pass data obtained by the method in this thesis can provide strong guidance for ecological evaluation and promote the improvement of fish pass design.Firstly,this thesis expounds the relevant research background and research significance,analyzes the research status at home and abroad,studies the classical image processing methods and template matching methods in the field of target recognition;summarizes the deep learning algorithms and image processing methods;and then determines the overall technology route.Secondly,this thesis constructs a fishway monitoring system for hydraulic engineering from the aspects of fish pass structure,system principle and key equipment selection.This thesis proposes a data set augmentation method combined with traditional image processing methods,which can achieve rapid data set expansion under the same detection.The experiments in this thesis show that under the condition of the same accuracy requirement,the data set augmentation method can reduce the number of model iterations and deploy the model more quickly.Thirdly,based on the research on the standard YOLOv3 network framework,this thesis has carried out a series of improvements and optimizations to improve the detection accuracy and detection speed of the standard YOLOv3 framework.For the detection speed problem,this thesis simplifies the network complexity and uses a 37-layer convolutional network to replace the original 53-layer convolutional network;for the problem of low recognition accuracy of small target objects,this thesis upgrades the original 3-scale detection to 4 types,so that the neural network pays more attention to shallow information,which is suitable for small target detection,and uses the K-means++algorithm to regenerate the Anchor value;for the relationship between network loss calculation and detection accuracy,GIOU is used instead of IOU to improve positioning accuracy;for target tracking matching and counting,using the Deep-Sort method,combined with the Kalman filter method and the Hungarian matching algorithm.The recognition accuracy rate of the network model obtained in this thesis is 94.6%,and the detection speed is 30.5FPS;and a fish size calculation method based on target recognition is proposed,which can effectively calculate the fish size.Finally,this paper uses PyQt5 to develop a visual interface,which connects the AI algorithm with the actual needs through the interface;lists the fish passing data information in the fish pass during the monitoring period and makes a brief analysis,which verifies the water conservancy and hydropower fish pass based on this thesis.Practicality and feasibility of fish identification and information statistics system.