Research And Implementation Of Water Conservancy Intelligent Video Monitoring System Based On Edge Computing

Water resources are the most important natural resources for the survival,life and production of human beings.How to realize the sustainable development of water resources has become an issue of great concern to the whole world.At present,my country’s water resources management methods are backward,water conservancy data acquisition is still mainly based on manual observation,the timeliness of the data is poor,the accuracy is low,and the intelligent management of water resources is imminent.The intelligent video monitoring system is an important part of the intelligent water management.Relying on a stable and efficient water conservancy intelligent video monitoring system,it can obtain real-time image and video data of the water conservancy monitoring site and perform automatic analysis and early warning.The system can save a lot of human resources and improve management efficiency.However,the generalization ability of the algorithm based on traditional machine vision to monitor water conservancy scenes is weak,and the model parameters need to be adjusted frequently in the face of complex water conservancy scenes with many uncertain factors,which is not popular.With the development of deep learning,water conservancy scene analysis algorithms based on deep neural networks have good detection results,but deep learning models are highly complex and have high requirements on the performance of deployed hardware devices,which are not suitable for low-power edge Calculated in the device.Therefore,this thesis studies and proposes a lightweight semantic segmentation network suitable for segmentation of different water conservancy scenes,and based on this network,proposes two water conservancy scene video analysis algorithms suitable for computing on Atlas 500 edge devices,and finally designs and implements Water conservancy intelligent video analysis system based on edge computing.The main work content is as follows:(1)A lightweight semantic segmentation network suitable for water conservancy scenes is proposed.After comparing the characteristics of the current semantic segmentation network,the ENet lightweight semantic segmentation network is selected,and its structure is redesigned,through the integration of mixed cavity convolution,high and low layer feature fusion,BN layer fusion,custom upsampling operator,Model conversion and quantification,etc.,researched and proposed a lightweight semantic segmentation network model Water＿ENet based on Ascend310 AI chip calculation and oriented to water conservancy scenes.The experimental results show that the Water＿ENet network can overcome the problem of the original network segmentation edge grid effect and the insufficient detection ability of small targets.Compared with the original network,the accuracy rate is improved by 3.01%.The model size is 0.81 MB after quantization,and the detection speed is on the Ascend310 AI chip.It is 32.42 ms,and the NPU has a small memory footprint,which can run in real-time in Atlas 500 edge devices.(2)Two new algorithms for water conservancy scenarios are proposed.1)A water level detection algorithm based on deep learning is proposed.First,the position of the water level is determined by the Water＿ENet network segmentation,and then the part of the water level image is detected through the Ruler＿net network designed in this thesis to obtain the water level value.Experiments show that the algorithm not only solves the shortcomings of traditional machine vision for water gauge scale extraction,but also can effectively overcome interference factors such as illumination and tilt.The accuracy reaches 93.51% and the speed is 68.4ms.2)Propose a detection algorithm for illegal laundry behavior in water conservancy scenarios.Firstly,people and clothing are separated through the Water＿ENet network,and then combined with the distance between the waterfront and the target to make a preliminary judgment,and the target that meets the condition is sent to the Mobile Net v2 classification network embedded in the SE attention mechanism module for classification.Experiments show that the algorithm can accurately detect laundry behavior with an accuracy of 94.91%and a speed of 48.2ms.(3)Design and implement a water conservancy intelligent video monitoring system based on Atlas 500 embedded edge devices.The system is divided into an edge computing layer,a center management layer,and a terminal display layer.The edge computing layer completes the deployment of water conservancy algorithms through the Matrix framework,and reasonably allocates the streaming,preprocessing,hard decoding,algorithm model reasoning,and post-processing engine modules to Atlas500 Host side processor and Device side AI chip calculation.The center management layer is mainly responsible for managing edge computing layer devices and algorithms,receiving analysis result information,and storing and forwarding video streams,and providing display data for the terminal display layer.The terminal display layer is responsible for providing users with video and analysis results display.The system has been deployed and operated in the Chengdu Water Conservancy Department and has certain practical value.