Research On Indoor Personnel Detection Technology Based On Panoramic Vision

Recently,with the continuous improvement of public safety awareness,and the rapid development of computer vision technology,target detection and tracking algorithms are often used in the fields such as mobile robots,autonomous driving,and large-scale monitoring systems.Although domestic and foreign researchers have got great results in the panoramic vision target detection area,there are still problems which haven't been solved.For example,accuracy and real-time detection cannot be taken into account at the same time.In this paper,to achieve great results on peson detection in indoor scenes,the target detection algorithm was combined with the target tracking.The main research contents are as follows:1)Aiming at the problem of image distortion using the fish-eye camera,the circle-based geometric model to correct the fisheye image was selected in this paper.In order to eliminate the problem of "short and coarsening" of the target in the expanded fisheye image,a preprocessing method based on horizontal down sampling was proposed.After the above two steps,the problem of "short and coarsening" of the target in the image is solved,and the preprocessed fisheye image is more suitable as the input of the target detection and tracking model.2)During the target detection period,the method based on YOLOv3 model with deep separable convolution structure,was prosed in this thesis.First of all,in order to make up for the time of the image preprocessing,the standard convolution structure was replaced with a deep separable convolution structure which has fewer parameters.Secondly,aiming at the inaccurate positioning of the YOLOv3 algorithm in the actual detection task,the bounding box loss function of YOLOv3 was reconstructed.In order to enable the model to achieve more accurate bounding box regression during the training process,Io U,a measure of the target detection effect,was combined with the corner distance of the bounding box,and a bounding box loss function based on A-Io U was designed.3)In the target tracking stage,aiming at the target ID conversion problem before and after the occlusion,the feature extraction network associated with the target appearance information during the tracking period,was improved.The Inception structure was combined with the Res Net network to build a multi-channel feature network structure,Res-Inception,and the residual module was replaced with the improved structure in the original feature extraction network.After improvement,feature extraction network could combine feature maps of different scales,enhance the expression ability of target features in the image,and reduce the distance deviation of target appearance features before and after occlusion.Finally,in the personnel detection task based on panoramic vision,the improved A-YOLOv3 algorithm was used as the detection module in the Deep Sort,which achieved high robustness and real-time performance.By increasing the Inception network structure and combining feature maps of different scales,the ability to express target features in the image was enhanced.The frequency of the same target ID transformation in the occlusion scene was reduced.Then,the improved A-YOLOv3 algorithm was used as the detection module in the Deep Sort target tracking process to achieve high robustness and real-time performance in personnel detection tasks based on panoramic vision.