| Video based traffic object detection and tracking technology is one of the research hotspots in the field of visual intelligent perception,and is a key link in tasks such as traffic element recognition and scene analysis.In actual traffic scenarios,factors such as weather conditions,light intensity,local occlusion,and appearance deformation all lead to changes in the appearance characteristics of traffic targets,thereby increasing the difficulty of traffic target detection.Especially in foggy scenes,the optical attenuation and scattering effects of the medium can cause significant distortion,degradation,and loss of feature details in the scene radiation;Target occlusion is also prone to occur in foggy weather,and frequent switching of target ID during the tracking process leads to a decrease in tracking stability;Meanwhile,deep learning based object detection methods rely on the training model of foggy image datasets,which are scarce and difficult to annotate,further increasing the difficulty of foggy traffic object detection and tracking.To address the above problems,this thesis relies on the key research and development plan project of Zhejiang Province,"Research on holographic digital perception technology for smart highways"(211424200512)to conduct in-depth research on video based traffic object detection and tracking methods under foggy conditions,and carry out relevant experimental verification.,and the innovative results achieved are as follows:(1)A foggy image level domain adaptive object detection method is proposed.The method firstly designs a differentiable foggy image enhancement module based on contrast,hue,color temperature and other features to reduce the impact of foggy image degradation on feature extraction;secondly adds an image-level adaptive component to the YOLOv4 backbone network based on domain adaptive techniques to improve the adaptive capability of the model in foggy weather;then designs two image-level domain adaptive strategies to enhance the use of different Finally,a differentiable image enhancement module is combined with the imagelevel domain adaptation method to form an end-to-end image-level domain adaptive object detection method for foggy weather.The method improves the migration capability of the object detection model under foggy weather conditions,and its superiority is verified by ablation experiments on publicly available datasets.(2)A foggy instance-level domain adaptive object detection method is proposed.Firstly,by adding a self-attention mechanism to the YOLOv4 feature fusion network to obtain imagelevel features;secondly,a corresponding domain adaptive component is designed for the instance-level features;finally,the above foggy sky image-level domain adaptive method is combined with the instance-level domain adaptive method to propose a complete foggy domain adaptive object detection method combining image-level and instance-level.The method solves the problem of negative migration caused by background interference in the detection process of the foggy image-level domain adaptive method,and further improves the generalization ability and detection performance of the model.(3)An improved Deep Sort object tracking method is proposed in this paper,aiming to solve the frequent ID switching problem caused by occlusion of traffic objects in foggy scenes.Based on the original Deep Sort algorithm,the Ghost module is introduced to improve the representation power of the original appearance feature extraction network.Furthermore,motion features are combined for data association matching to improve the overall accuracy of the object tracking algorithm.(4)A set of experiments was designed to evaluate the proposed video-based traffic object detection and tracking method under foggy conditions.Experimental results on public datasets demonstrate the superiority and robustness of the proposed method in foggy object detection tasks.Compared with networks that do not use domain adaptation methods,the mean average precision(m AP)of object detection is improved by about 10%.Compared with two-stage fog domain adaptive methods,the detection speed is increased by 40%.Regarding object tracking,the proposed method reduces trajectory switching frequency by 15% and improves tracking accuracy by 4%.In real foggy highway video object detection and tracking experiments,the proposed method stably tracks objects,and algorithmic real-time performance is also ensured. |
