Pedestrian Attributes Recognition Based On The Deep Learning In Outdoor Scenes

In recent years,video surveillance system has been widely used in daily life.In the high-speed junctions,crossroads,parking lots and other places,a large number of surveillance cameras are installed and deployed,which provides a large number of video surveillance data.People are eager to get more useful information about the pedestrians in the video,such as gender,age,appearance and so on.In the traditional monitoring system,pedestrian attribute information is obtained through human observation,which is inefficient and can not meet the practical application requirements.For the traffic monitoring scene,this thesis proposes a multi-depth feature fusion pedestrian attribute recognition algorithm.The algorithm is mainly divided into two stages:deep feature extraction and deep feature fusion.In the stage of deep feature extraction,we propose a multi-task convolutional neural networks.The convolutional neural network can input different attributes and simultaneously extract pedestrian's gender,age and travel style features.The CNN that we trained share convolutional layer between the attributes,not only make full use of the link between the attributes,reducing the number of network parameters and improve the efficiency of feature extraction.In the stage of feature fusion,we propose a decision-level feature fusion method based on recognition results,and further improve the recognition rate of pedestrian attributes compared with the feature-based fusion method.In addition,we establishe a dataset of pedestrian attributes under traffic monitoring scenarios and annotate the data.Different from the previous datasets,this dataset is mainly for traffic scenes,and is mainly aimed pedestrian pictures with lower resolution.The dataset includes a total of 21,385 images,all from real traffic surveillance videos,with diverse backgrounds.