| In view of the traditional crowed abnormal behavior detection real-time is not high,Moving target detection is not adaptive,abnormal behavior recognition rate is low.According to the research results of the theory and the latest domestic and overseas,moving target detection using adaptive Vibe algorithm,an adaptive segmentation algorithm based on an accelerated segmentation algorithm and an extreme learning machine classification algorithm are used to detect the abnormal behavior of small and medium sized crowd.At the same time,the GKLT algorithm and the ensemble detection algorithm are used to analyze and simulate the high density group behavior.Detailed works are as follows.Firstly,Improved adaptive Vibe algorithm.There is not much difference between the improved Vibe algorithm with traditional Vibe algorithm in background detection and model updating.In this paper,we focus on improving the process of moving target recognition.The traditional Vibe algorithm uses fixed threshold to determine the moving target pixels,The fixed threshold will not increase or decrease appropriately for environmental changes,this will lead to the complex environment of the traditional Vibe algorithm detection results are very vague,simple environment is better.To solve the above problems by using the adaptive threshold determination,the minimum Euclidean distance between the pixel in the background model and the current pixel is recorded.After processing the minimum Euclidean distance,the fixed threshold R will increase or decrease with the complexity of the environment.Finally,the improved motion detection algorithm is compared with the codebook model and the traditional Vibe model.Secondly,Adaptive generalized corner algorithm.FAST corner extraction algorithm uses ID3 greedy algorithm to construct decision tree,this may lead to a decision tree to local optimum.In addition,FAST corner extraction algorithm uses a trinomial tree,but for the corner extraction of target is two yuan.In view of the above problems,we use a combination of two trees instead of the FAST corner algorithm’s trinomial tree.Thirdly,Extreme learning machine classification algorithm.Aiming at the problem of low recognition rate and poor real-time performance.In this paper,four kinds of features are extracted and classified by the learning machine algorithm and compared with the hidden Markoff classification algorithm and SVM classification algorithm.The experimental results show that the recognition rate of ELM algorithm is higher than that of the classification method in the literature.Fourthly,High density population behavior analysis.The GKLT algorithm is used to track the high-density crowd,and the position and velocity of the feature points are recorded.Combined with the degree of convergence detection algorithm,we can detect which groups of people in the direction of movement are consistent and expressed in the same color.For groups of different directions,different colors are used.The algorithm can effectively help the monitoring personnel intuitively to see the same direction of movement of the population cluster in high density group,According to the characteristic point velocity extracted by GKLT,the average velocity of the group motion can be directly reflected in the monitoring center.It will be more convenient for regulators to understand the motion information monitoring range of population. |
