Research On The Recognition Model Of Abnormal Wandering Police Events

In recent years,China has paid increasing attention to the construction of social security prevention and control system,especially in the field of infrastructure construction and equipment development has made great progress.However,there are still deficiencies in the monitoring and early warning of social security incidents,which seriously restricts the orderly advancement of intelligent security community.This topic comes from "Key Technology and Equipment for Intelligent Monitoring and Identification of Police Incidents Based on Multi-information Fusion in Public Security Prevention and Control Places" of the national key research and development plan project "Key Technology and Equipment for Intelligent Detection and Warning of Social Security Incidents",focusing on the identification of abnormal prowling alarm incidents.This paper mainly uses Pedestrian re-identification,Clustering technique,three decision-making models,combined with the characteristics of pedestrian tracking data,building a recognition framework of abnormal wandering alarm events,to realize the recognition of abnormal wandering alarm events in the video.Aiming at the problem of broken and drifting pedestrian tracking trajectory,a pyramid pedestrian re-identification model was constructed to obtain a complete trajectory in the pedestrian field of view.Aiming at the problem that the performance of pedestrian re-identification methods is greatly reduced in practical application scenarios due to the domain differences,an adaptive metric learning model based on negative samples is proposed,and the iterative strategy of the metric learning model in practical application scenarios is given.Experimental verification on Market-1501 and Duke MTMC-re ID dataset shows: the performance of pyramid pedestrian re-identification model is better than the existing Res Net deep network algorithms such as Aligned Re ID and Partloss,and the adaptive metric learning model effectively improves the domain difference problem of the recognition model.In order to solve the distortion problem of the existing abnormal wandering behavior recognition methods,the research proposed a theme-based abnormal wandering behavior recognition clustering algorithm.The algorithm first segmented the pedestrian trajectory according to the direction of motion,and then performed POI-based trajectory theme mining,used the Sparse DTW algorithm to calculate the similarity of sub-trajectories within the same trajectory topic,and then used the density clustering DBSCAN algorithm to complete trajectory clustering and identification Abnormal trajectory.The experimental results on the CAVIAR dataset show that compared with the K-Means clustering algorithm and single DBSCAN algorithm,this algorithm is more accurate in identifying abnormal trajectories,and the clustering results are more practical.Aiming at the problem that the existing sequential three-branch decision model lacks a comprehensive measurement of the overall loss cost,a cost-sensitive sequential three-branch decision model is proposed to comprehensively measure the test cost and misclassification cost.Three decision-making identification models for abnormal wandering alarm events are constructed.In order to achieve the description of the boundary domain objects in the model,the research proposed a probabilistic description algorithm based on the KNN method.Based on the Abnormal Wandering experiment dataset,the cost-sensitive sequential three-branch decision model was validated.The above research results have enriched the research content of the subject "Key Technologies and Equipment for Intelligent Monitoring and Recognition of Police Events Based on Multi-information Fusion in Public Security Prevention and Control Sites".The specific algorithms have been implemented and applied and integrated into the subject in the system platform.