Pedestrian Attribute Recognition Based On Reinforcement Learning

The rapid development of economy and technology has made the society change rapidly from industrialization to informatization.Modern urban monitoring systems have also ushered in rapid development,and multimedia data such as images,audios,and videos are increasing day by day.Efficient processing of massive multimedia data has become a problem that cannot be ignored.Pedestrian attribute recognition is a very important research content in this area.It can obtain the attributes of pedestrians based on pedestrian images,which plays an important role in analyzing pedestrians and tracking criminals with potential suspects,and is of great significance to the construction of a safe city and a harmonious society.Most of the existing methods use deep learning technology to take the recognize pedestrian attribute recognition as a recognition task.In fact,the process of pedestrian attribute recognition can also be regarded as a decision-making process.Therefore,in this dissertation,the pedestrian attribute recognition task is regarded as a decision task,and reinforcement learning technology is used to recognize the pedestrian image attributes.Specifically,in order to further optimize the strategy of reinforcement learning agent,the reward function and loss function are innovated.The key work is mainly as follows:The dissertation first proposes a reinforced pedestrian attribute recognition algorithm based on group optimization reward(Rein-PAR).Firstly,the pedestrian attribute recognition is defined as a Markov decision process,the state space is constructed by image features and semantic features,and the action space is constructed by 0 and 1,and the corresponding state transition process and reward function are designed.Then the pedestrian attributes are grouped according to their regions and characteristics to alleviate the adverse effects of inter-attribute imbalanced data distribution.Then,a group optimization reward function is proposed to deal with the imbalanced data distribution of intra-attribute to obtain a better strategy.Finally,Deep Q-learning algorithm is used to train the agent to get the strategy.In addition,this dissertation also proposes a reinforced single attribute recognition algorithm for pedestrian attribute recognition(RSAR-PAR),the main purpose is to more effectively deal with the problem of imbalanced data distribution of intra-attribute.Firstly,the algorithm defines the pedestrian attribute recognition problem as a process of recognizing multiple attributes separately.Then,an appropriate Markov decision process is formulated based on the definition,and reward function is used to alleviate the adverse effects of the imbalanced data distribution of intra-attribute.Then a loss function combining reinforcement learning and supervised learning is proposed to better train the model.Finally,a suitable training process is designed to train agent.The comparative experiments with other methods on three public pedestrian attribute datasets and the specific recognition results prove the advancedness and effectiveness of Rein-PAR and RSAR-PAR.