Research And Implementation Of Pedestrian Recognition Technology Based On Adaptive Feature Matching

In recent years,with the rapid development of video surveillance,the demand for public security agencies to use video surveillance to find criminal suspects is also increasing.Intelligent video surveillance will also be the future development trend of video surveillance,and pedestrian recognition technology is intelligent security.One of the important applications.Due to problems such as blurred pedestrian images,changeable postures,and too small targets in the surveillance scene,the accuracy of pedestrian recognition under video surveillance is low.The current recognition method is based on the characteristics of the entire pedestrian picture.The method is too simple.It cannnot be flexibly used to identify the most valuable features of pedestrians,nor can it customize some attributes of pedestrians to achieve coarse classification of pedestrians.Therefore,the current pedestrians The identification method has strong limitations.The information available for pedestrians obtained in practical applications is different.When there is a clear retrieval of the face photo of the target,the facial features should be used as the unique identifier for identification.If you want to find men wearing glasses and short sleeves,you should obtain attributes Feature information for identification.Fully utilizing the multi-attribute features of pedestrians for recognition will be the mainstream technology for pedestrian recognition in surveillance,and the pedestrian recognition technology for adaptive feature selection is also imminent.In response to the above problems,this thesis proposes adaptive feature matching pedestrian recognition technology,the main work is as follows:(1)Research on anti-interference pedestrian detector based on YOLOv3-Ma.In view of the detection interference factors in the actual scene and the massive data frames generated by the monitoring system,this thesis analyzes the original model data initialization,there are problems such as the random selection of the initial cluster center and the interference of outlier noise data,and the initial cluster The central rule selection mechanism and the outlier noise factor removal mechanism make the pedestrian detector's test accuracy increase by about 1.5%in various test scenarios.In the face of massive monitoring data,if each frame of data is detected,the detection efficiency is greatly reduced.In this thesis,a detection mechanism for removing similar frames is added to effectively solve the problem of similar frame detection in the monitoring system,and the detection speed is obtained.Significant improvement.(2)Study pedestrian recognition technology for adaptive feature selection under video surveillance.In view of the problems of low recognition accuracy and single recognition method caused by the diversity of pedestrian poses in the monitoring system,this thesis first proposes a pedestrian attribute feature classification mechanism.According to different rough poses of pedestrians,different attribute features have different use values,and A mechanism for adaptively selecting feature recognition is designed.It effectively solves the limitation of relying on the identification of a single pedestrian attribute,and also realizes the coarse classification and recognition of pedestrians based on attributes,which meets the diverse needs of actual monitoring.At the same time,a quality judgment mechanism based on the Laplacian operator and the number of pixels is added to remove the input low-quality pedestrian data.In the end,the average accuracy rate under the test of a variety of complex scenes reached about 72%,which realized the feature refinement and coarse classification of pedestrians.In this thesis,after implementing the adaptive feature matching pedestrian recognition mechanism,this research is applied to actual video surveillance to meet the needs of current smart video surveillance such as real-time,accuracy,and diversity of recognition methods.