Pedestrian Attribute Recognition Based On Deep Learning

Pedestrian attribute recognition can turn pedestrian information in surveillance videos into high-level semantic information that can be used to search.It can assist in tasks such as pedestrian retrieval and pedestrian re-identification.Therefore,it has attracted more and more researchers' attention.Early pedestrian attribute recognition algorithms used manual designs rules to extract features for pedestrian attribute recognition,but the actual scene complexity,so it was difficult to obtain better recognition results.Therefore,this thesis proposes two improved algorithms based on deep learning:(1)In order to reduce background interference on attribute recognition,a pedestrian attribute recognition algorithm based on suppression of background interference is proposed.The algorithm can improve the attribute recognition effect,because it forces the features extracted to focus on the body region in the image.First,two branches are added.The backbone network extracts pedestrian image features.The added branches are used to separate the pedestrian body area from the background area on the feature map.Therefore,we can obtain the feature vectors from different areas.Then,the loss predicted by the network is calculated by the weighted cross entropy loss function,the area contrast loss function is used to calculate the loss caused by the wrong feature extraction due to background interference.Finally,the two part losses are accumulated as the total loss of the network,and update the parameters of the network.(2)In order to improve the recognition effect by locating the position of each attribute,a pedestrian attribute recognition algorithm based on attribute locating is proposed.First,the algorithm designed an attribute location module,which improves the network's attribute location capability by weighting each attribute at the regional level.Then,the attribute location module is applied to multiple levels of the network to form a multi-branch network,the network can obtain multi-level features containing location information.Finally,the multi-level features are linearly fused.The fused features include both the multi-level features and the position information of the attributes.The feature vector is used for attribute prediction to obtain the recognition result.Through the experimental verification two pedestrian attribute data sets.The two improved algorithms in this thesis have been improved by mean accuracy,accuracy,precision and other performance indicators compare with other existing method,which prove the effectiveness of the method.