Research On Infrared Pedestrian Target Detection Technology Based On Improved YOLOv3

YOLOv3 algorithm,as the most popular deep learning-based target detection algorithm,has the advantages of high detection accuracy and fast detection speed.However,the current deep learning-based target detection algorithms are mainly for targets in visible images,and when they are applied to infrared target detection,the detection performance of the algorithms is degraded due to the lack of information such as color and texture in infrared images and the susceptibility to complex background interference.In order to investigate the problem of low pedestrian target recognition rate after applying YOLOv3 algorithm to IR pedestrian target detection,this thesis improves YOLOv3 algorithm to make it have higher recognition rate for pedestrian targets in IR scenes.Based on the above problems,this thesis makes improvements to the YOLOv3 target detection algorithm,and the main work is as follows:(1)To address the problem that the a priori frame in the YOLOv3 algorithm is obtained by clustering the COCO visible dataset,which is not quite suitable for the infrared pedestrian target dataset in this thesis,the target frame in the dataset is re-clustered and analyzed to obtain a modified a priori frame,which is more suitable than the original a priori frame of the YOLOv3 algorithm.The modified prior frame is more suitable than the original prior frame of YOLOv3 algorithm,which is helpful to improve the recognition rate of the algorithm for infrared pedestrian targets.(2)To address the problem that the existing infrared pedestrian target dataset is small,data augmentation is performed on the existing dataset.Using data augmentation techniques can expand the dataset to a certain extent and also improve the robustness of the algorithm.(3)To address the problem that the feature extraction network of YOLOv3 algorithm is difficult to extract the features of pedestrian targets in infrared scenes,CSPNet is introduced into the feature extraction network of YOLOv3 algorithm,which increases the number of layers of the feature extraction network,divides the input into two parts for feature extraction,and then fuses the features of different sensory fields,which can better extract the features of infrared pedestrian targets.It further improves the recognition rate of the algorithm for infrared pedestrian targets,and also reduces the amount of network operations and the number of parameters of the model to maintain a high recognition rate while still maintaining a fast detection speed.In this thesis,we perform data enhancement on the publicly available OTCBVS dataset and experimentally verify the detection effect of the improved algorithm on the expanded dataset.The experimental results show that the improved target detection algorithm improves the m AP by 11.26% over the YOLOv3 target detection algorithm,while the detection speed is basically the same as that of the YOLOv3 target detection algorithm,and the improved algorithm obtained in this thesis is more The improved algorithm is more suitable for pedestrian target detection in infrared scenes.