| Pedestrian detection and tracking under occlusion has always been an admittedly difficult problem in the field of computer vision.In the highway outfield environment,due to the effects of lighting variations,varying camera heights,and numerous occlusions,pedestrians are characterized by multi-scale,variable scale,and diverse morphology,which increases the difficulty of pedestrian detection and tracking.Therefore,it is of great theoretical and practical significance to study pedestrian detection and tracking algorithms based on local occlusion in high-speed external field environments.Based on the in-depth study of YOLO algorithm for occlusion detection,the thesis constructs a pedestrian detection algorithm framework FFSA-YOLO applicable to highspeed outfield,and proposes an improved method for state classification,centroid relocation and auxiliary feature determination to address the problems of false tracking,ID jumping,localization jitter and target mismatch when Deep Sort is applied to the scenario in this thesis.The main work and contributions of the thesis are as follows:(1)For the problem of multi-scale pedestrian detection under occlusion,the thesis proposes a cross-scale feature fusion strategy in terms of feature pyramid and perceptual field,and designs a parallel shallow network to fuse high-level semantic features to enhance the local detail features of small-scale pedestrians;to further enhance small-scale pedestrian detection,a P2 layer is added and an extended perceptual field module based on hole convolution is fused to enhance the semantic information of the P2 layer Finally,to avoid the problem of uneven distribution of pedestrians with small scale span,a 2*2Deep Parts structure is designed for feature weighting and fusion detection to improve the overall pedestrian detection effect.(2)To address the problem of inconspicuous and missing pedestrian features under occlusion,we first conduct targeted data augmentation from the perspective of data enhancement and design the attention mechanism of multi-information fusion to enrich the contextual information of pedestrian features in the network and improve the fast feature extraction capability of the network for pedestrian targets in the occlusion environment;to solve the problem of poor detection due to the overlap between pedestrians,we introduce the Manhattan distance-based Confluence algorithm,which performs confidence-weighted statistics on the distance between mutual targets,is introduced to improve the detection accuracy of mutually occluded targets.(3)For Deep Sort’s false tracking and ID jumping problems,this thesis designs a secondary matching method based on the confidence rate of change to improve the false detection problem based on Io U matching,and proposes the classification processing idea of state judgment,using LBP features to determine the deformation and then separate the occlusion state.For the problem of positioning jitter caused by short-time occlusion,a center repositioning method is proposed to introduce deep features for position correction judgment of the center point to improve the prediction effect in the tracking process;for the problem of long-time occlusion,an improved Res Net-based pedestrian attribute recognition network is designed to assist Deep Sort in extracting deep features for target matching to improve the tracking accuracy of long-time occlusion;finally,in the video input stage and the stage of matching to predict the next frame to perform data dimensionality reduction to improve the speed of the tracking algorithm.The algorithm framework for partial occlusion pedestrian detection and tracking in the highway outfield environment is constructed by synthesizing the above.By collecting high-speed outfield surveillance videos for testing,experiments show that the algorithm in this thesis has improved in speed and accuracy,and experiments on Caltech public dataset demonstrate the better advantage of the pedestrian detection algorithm in this thesis. |
PDF Full Text Request