Research On Pedestrian Detection And Attribute Recognition In Surveillance Scenarios

In recent years,with the construction of smart cities and safe cities,monitoring systems have been widely used in public places.It is helpful to reduce the burden of artificial observation and improve the efficiency by giving the intelligent ability of the monitoring system.The technology of pedestrian detection and pedestrian semantic attribute recognition under surveillance system has become the research hotspot in the field of surveillance video.Pedestrian detection is the foundation of intelligent monitoring system,which aims at locating pedestrians from images,and requires high detection rate and realtime performance.The semantic attributes of pedestrian,such as gender,age,type of clothing has a broad application prospect.It can bridge low-level features and highlevel human cognitions,which help to design more intelligent human-computer interaction systems and effectively improve the performance of pedestrian tracking and re-identification technology.However,in the surveillance scene,pedestrian attribute recognition still has challenges,because pedestrian in the picture usually account for a small part with low resolution and there exist some uncontrollable interference,such as illumination and camera viewing angle.In view of the above problems,pedestrian detection and attribute recognition in surveillance video are studied in this paper.The main contents include:In order to improve the real-time performance of pedestrian detection,ViBe algorithm is used to build the background model.For the case of fracture and void in the extracted pedestrian foreground,the vertical rectangular window statistical filtering method is used to replace morphological treatment,which can effectively connect the broken part of the pedestrian.Aiming at the problem that the dimension of color self similarity is too large,a symmetric color self similarity feature is designed and combined with HOG features.Using SVM(Support Vector Machine)to detect pedestrians in the foreground area can reduce the search area,improve the efficiency,and reduce the false detection rate.In the aspect of pedestrian attribute recognition,faced with complex and diverse attributes,it is difficult to extract effective feature descriptions for classification based on existing feature operators.With the development of deep learning,multi-layer convolutional network has been proved to have excellent feature extraction ability.In this paper,the deep convolution model is used to realize attribute recognition,and a LSTM model based on attention is designed on the basic network.Comparing experiments on the data set,the model designed in this paper has better results.