Moving Object Detection And Human Behavior Analysis In Surveillance Video

The moving object detection and human behavior analysis technology based on surveillance video is the key technology of the automatically human monitoring system.The technology is widely used in many filds including traffic supervision,military aerospace and sports,especially for safety precautions in public places where there are a lot of people such as airports and railway stations.Artificial supervision takes a long time,the workload is large and easy to leak.Therefore,it is of great significance and value for detection of moving targets and recognition of human behavior automatically based on surveillance video.This thesis consists of three aspects: including moving object detection,multi-pedestrian tracking and human behavior analysis.In this thesis,the moving target area in the original video is detected with a moving object detection method by fusing texture and color features with confidence,and the human body in the moving target is detected and tracked.Finally,the five kinds of behaviors including walking,running,jogging,clapping and waving are classified based on the human behavior key frame sequence obtained by tracking.The main contents of this thesis are as follows:(1)Comparing the advantages and disadvantages of the color features and texture features in the detection of moving objects,a moving object detection method by fusing texture and color features with confidence is proposed to solve the disadvantages.This method constructs the background model together with the scale invariant local ternary pattern(SILTP),RGB color information values and the respective confidence levels,and then classifies the pixels according to the similarity degree matching.The method has certain robustness to shadow and light changes,and can deal with complex dynamic background effectively.(2)DPM(Deformable Parts Model)is a multi-model and multi-scale pedestrian detection algorithm with the highest level of pedestrian detection.Based on DPM's good pedestrian detection algorithm,a multi-pedestrian tracking framework is proposed.Based on DPM's good performance on detection,a multi-pedestrian tracking framework is proposed in this thesis.The framework tracks each pedestrian based on the single-person tracker of the Kalman-fittered,then uses JPDA data association algorithm to correlate the detection with the target to achieve multi-pedestrian tracking.The tracking framework has a good tracking effect on multi-scale pedestrian.(3)The traditional method of human behavior analysis based on the temporal and spatial features of the local features is improved.The traditional algorithm classifies the behaviors by extracting the temporal and spatial features of the human motion,and calculates the similarity matching on the clustering obtained the K-means.This thesis adds the features of the aspect radio and speed of the tracking people.Experiments show that the improved algorithm performs better to distinguish walking,running and jogging from each other,the addition of the aspect radio feature also increases the recognition rate of clapping and waving.