Research And Implementation Of Pedestrian Detection Algorithm In Unanchored Thermal Infrared Images

Pedestrian detection is a hot area of research in the field of computer vision target detection,and thermal infrared image pedestrian detection techniques are receiving increasing attention in intelligent assisted driving and video surveillance scenarios as they are not restricted by weather and time.In this paper,the anchor-free target detection algorithm FCOS is improved based on thermal infrared images,and the main research content and innovations are as follows:First,a depth saliency map-based thermal infrared image enhancement technique is proposed for the enhancement of pedestrian targets in thermal infrared images.By annotating some pedestrian instances of thermal infrared images in the KAIST dataset,a depth saliency network is trained to extract the saliency map,and then the saliency map extracted from the thermal infrared images containing only pedestrian targets is fused with the original thermal infrared images in a pixel-level weighting,the enhanced thermal infrared images improve the contrast and highlight the pedestrian targets in the images,which helps the detection algorithm to extract pedestrian features.Second,to address the problem of high false detection and missed detection rates of thermal infrared image pedestrian detectors,the following improvements are made to the anchor-free FCOS target detection algorithm: by introducing the channel attention module SE-Block into the Back Bone network of the FCOS algorithm,which is used to learn to model the relative importance of different feature channels,the weights of the features extracted by the convolutional neural network are reweighted and calibrated to improve the weights that play a greater role in pedestrian target detection.In the post-processing part of the algorithm,the original non-maximum suppression NMS is improved to soft threshold non-maximum suppression Soft-NMS,which is used to reduce the probability of missed detection of obscured pedestrians.The experimental results show that the detection effect of this method has been greatly improved compared with the original algorithm,which effectively improves the problem of missed and false detection of pedestrian targets in thermal infrared images based on the original algorithm,and at the same time can meet the demand for real-time performance in realistic scenes.Final,the trained model is used to build a web browser-based platform for thermal infrared image pedestrian detection,which enables users to upload thermal infrared images or video files to detect pedestrian targets in the screen.