Appearance And Gait Features Based Pedestrian Identity Recognition

With the continuous development and progress of society,the coverage of monitoring equipment in the city is higher and higher.The pedestrian appearance information(including face,clothing,body shape)and walking gait information have high reliability and stability in a certain time limit.Therefore,it is very suitable for people in public places or stores that pay attention to customer experience to complete the identification work with the help of people’s appearance,gait and other information.In this thesis,firstly,the appearance feature extraction network and gait feature extraction network of pedestrians are improved.Then we try to combine the appearance features and gait features to get a more robust pedestrian identification system.In the aspect of appearance network,we use a multi-scale human body feature extraction network.From the perspective of loss function of measurement learning,this thesis proposes to introduce the topological relationship between pictures in mini batch into the existing triplet loss function,that is,to consider that two pictures belonging to the same category should have similar topological relationship network in the feature space.Two improved loss functions are proposed.Compared with the basic algorithm,the m AP accuracy of the proposed loss function on the two open datasets of markt1501 and duke MTMC-Reid is improved by3.75 % and 5.69 %,respectively.It has also achieved good results in other pedestrian recognition data sets.However,the appearance information of pedestrians is sometimes confusing.Therefore,we have a study on gait recognition.In this thesis,we propose an angle based "divide and conquer" algorithm for gait recognition.Firstly,the probability information of different shooting angles of gait sequence is obtained through multi task learning.On this basis,according to the probability of different angles,different network parameters are used to learn the characteristics of corresponding angles.The final feature includes not only the general feature which is responsible for cross angle information extraction,but also the angle feature which is used to extract the common information of the image in the adjacent angle.According to the previous probability,the network parameters of different angles are optimized in different degrees.In the open gait data set,the performance of the original algorithm is improved obviously under some test conditions.In order to discuss the function and application difficulties of appearance information and gait information in the actual monitoring system.From the reality,we collected a small-scale actual monitoring data set,and completed the complete process of pedestrian contour extraction,image preprocessing,appearance feature and gait feature extraction,as well as the integration of appearance and gait features under the conditions of open data set and actual monitoring data set.The experimental results show that the comprehensive use of appearance and gait information plays a more powerful role in the identification of pedestrians under some test conditions.To some extent,it can provide some reference for the comprehensive practical application of appearance information and gait information.However,it needs to be acknowledged that there are still some problems in the actual application of the current model under the actual data conditions.We also put forward some thoughts and suggestions on the current problems and the difficulties to be solved in the future.