Research On Video Recognition Of Urban Management Case Based On Deep Learning

With the acceleration of urbanization,and the growth of urban population,strengthening urban management,and promoting the construction of smart cities are receiving attention by governments at all levels.At present,the domestic cities have formed a large personnel of the city Management Brigade,spending a large amount of money to buy fixed surveillance cameras.However,fixed surveillance cameras have limited perspective,making it difficult to monitor the high-risk areas of urban management cases such as backstreet lane.Meanwhile,watching surveillance video manually is not only time-consuming,but also inefficient.In response to the practical problem,this thesis uses the real case pictures as the training basis to propose a target recognition algorithm of urban management case which can be deployed on mobile phones or low performance devices by combining deep learning knowledge.In order to achieve the goal of flattening the handling of urban management cases and improving the management efficiency and intelligence degree,this algorithm can identify the target of the case taken by the mobile camera,alleviate the defects with limited perspective of fixed camera,accelerate the classification and allocation of cases.The main work was as follows:1.According to the actual needs of city officials and the relevant state regulations,this thesis uses the real case pictures accumulated in the data sets of the city management system commissioned by Qingxiu Distract and Jiangnan Distract of Nanning City to construct a new data set named "City Management".This data set marks 6646 pictures in a manually labeled way,including 8 classes.2.Based on the "City Management" data set,an improved SSD target recognition algorithm is proposed by combining the improved MobileNet with SSD target recognition algorithm to solve the problem that the existing fixed-mounted surveillance cameras have monitoring dead angles and low hardware performance of mobile devices.The accurate detection of the common eight specific urban management case targets was achieved by the proposed algorithm which uses the camera provided by mobile device to capture the scene video.The mean average accuracy of the detection was 15.5 percentage points and 10.4 percentage points higher than the prototype YOLO and the prototype SSD,respectively.3.On the basis of the improved target recognition algorithm,this chapter combined the recognition results with SURF algorithm,FLANN algorithm to realize the function of case identification and monitoring,such as fire warning,suspect identification,human flow monitoring and so on,which can maximize the use of monitoring video,improve management eff-iciency and reduce reaction time.4.This thesis integrates the recognition algorithm proposed above into a city management APP.This app can be deployed on the mobile side of the IOS mobile phone,which would use the mobile phone with its own camera to shoot case pictures or crime scene video.Then the algorithm deployed on the mobile phone can realize the identification and classification of case objectives.The grid personnel management subsystem will decentralize the case reported by the app to the specific responsible person,which can improve the city management efficiency.