| With the Ministry of Public Security "Golden Shield Project" and the national "safe city" development of the construction,the city's video surveillance system was built in most cities.At present,the city surveillance is still in the stage of "record no judgment".With the geometric increase of the video data volume,the traditional manual judgment is time-consuming and inefficient.How to automatically detect pedestrians and recognize action in the large number of video data is an urgent problem to be solved.This paper focuses on pedestrian detection,action feature extraction and action recognition in urban intelligent surveillance system.The main contents are as follows.First,pedestrian was detected from urban surveillance video.Through the experiment,the frame difference method,the optical flow method and the background subtraction method were analyzed.As a result,the mixed Gaussian background modeling method was selected to extract the moving target.And then machine learning and deep learning were used to detect pedestrians from those moving targets.In the case of machine learning,the AdaBoost classification method based on Edgelet feature and the SVM classification method based on HOG feature were used for the extraction and recognition of the human feature,respectively.Our research shows that the latter performed better in moving pedestrian target detection.In regard to deep learning,a pedestrian detection algorithm based on the improved YOLO network model was proposed.By using the INRIA pedestrian database,the method achieved 93.1% accuracy in the experiment,which could well detect the pedestrian in the monitoring scene.Secondly,basing on the detection of pedestrian,the algorithm of human action feature extraction was studied.The static,dynamic and spatial-temporal feature analysis indicates that the latest feature which integrates dynamic and static features is the most suitable descriptor of human action.Then two spatial-temporal features were studied,one is the MHI feature,the MHI feature based on regional constraints was proposed in this paper.The other is the spatio-temporal interest points.This paper presented an integrated feature of 3DHOG and HOF.Finally,two kinds of action recognition methods based on template matching and theme model were adopted to realize the core function of urban intelligent surveillance system.For the template matching,the MHI template is very popular,but not accurate enough,resulting in a bad classification effect,so an improved algorithm of template matching based on region constraint MHI was proposed.For the theme model,an improved YOLO network with the 3DHOG and HOF fusion features was used to solve the background noise.The theme model only focuses on spatio-temporal interest points around the pedestrian,which improves the robustness.In the KTH behavior database,the accuracy rate reached 92.7%,while in the more difficult CASIA database it only reached 84.7%.In this paper,the pedestrian detection and action recognition method for urban intelligent surveillance were systematically studied,and the above results were applied to the specific database.The experiment shows the effectiveness of the algorithm and the achievement of the requirements for the pedestrian detection and action recognition.Our research has important reference value for the study on urban intelligent surveillance such as complex scene and pedestrian. |
