Research On Human Head Detection In Real Time Video

With the constant economic and social renewal in China,remarkable achievements have been made in the construction of urbanization,and the social contradictions such as public security risks have also been aggravated while improving people's daily life.The rapid growth of urban population caused by urbanization leads to the sudden increase of human flow or population density in the unit area,and the public safety incidents are common.As an effective technology of security monitoring,head detection in real-time video plays a key role in the statistical analysis of human flow or crowd density.This paper analyzes the head detection technology,and summarizes the previous research results of learning,improvement and optimization of some deficiencies,to achieve accurate detection of pedestrian head in the video scene,and provide a more reliable detection method in human traffic statistics analysis and other practical applications.The detailed work is as follows:1.The Kalman filter modeling method is used to establish the background model of pedestrian movement in this paper.The modeling method for Kalman Gain model is used to solve the problem of pedestrian loss when modeling the movement state of pedestrian movement without irregular movement,and the improvement measures of Kg dynamic update are put forward,the possibility of the loss of the target caused by the diversity of the pedestrian movement state.2.The HOG features can be well expressed for the distribution of the gradient or the edge of the extracted region.However,the feature itself does not have the scale invariance.In practical applications,only the target object with the same size as the sample picture can be detected,proposed a human head detection method based on multi-scale collaboration.The advantage of linear support vector machines is exploited for classification decision making and the extracted HOG feature are trained offline by a cooperative classifier.In real-time target detection stage,the input video frame sequence is analysed at multi-scale to obtain the frames to be detected under different resolutions.The human head detection is performed at different scales and the results are stored.After that,the detection results at each scale are fused and calibrated.With this process to improve accuracy and detection efficiency.3.In the multi-scale coordination head detection system,the traditional gradient direction histogram is applied to the field of high-definition video surveillance,which cannot meet the real-time requirement of monitoring video due to the massive calculation of feature extraction.This paper proposes a parallel head detection method for accelerating GPU_CPU heterogeneous parallel computing based on GPU HOG server is responsible for feature extraction of large intensive block,and the CPU is responsible for testing the implementation of other modules in the process.This heterogeneous parallel poll method can give full play to the advantages of CPU and GPU.4.The traditional parallel reduction algorithm for HOG feature extraction and the time complexity is not ideal,proposed an improved parallel reduction calculation method,through the "parallel scan" method,to reduce the number of nodes is calculated,HOG reduces the time complexity of feature extraction,but also can satisfy the multi-resolution head detection of real-time requirements.Experiments show that the method proposed in this paper is superior to the traditional methods in the detection rate,recall rate,accuracy and detection rate of the pedestrian in the same surveillance video.