High-precision Detection And Cross-camera Motion Reconstruction Of Traffic Objects

In recent years,the idea of obtaining extensive and in-depth traffic information through massive traffic surveillance video data and integrating with new infrastructure has increasingly become a new focus of transportation.However,in the current research,the dedicated public datasets for traffic objects with rich information are still incomplete,and the excellent detection of objects with dramatic scale changes and the accurate positioning of road spills are still difficult problems.In addition,the incomplete extraction of traffic object features leads to discontinuous acquisition of object motion trajectories,which makes it impossible to reconstruct and analyze the cross-camera motion process of traffic objects,resulting in a waste of monitoring resources.This paper focuses on the high-precision detection and cross-camera motion reconstruction of traffic objects.The main research contents are as follows:(1)A multi-category object dataset for highways is constructed.The traffic object dataset from the monitoring perspective covers 10 traffic object categories,more than 900,000 sample annotations of different sizes and poses,and a total of more than 300,000 images of different weather and lighting conditions.Through image cropping,insertion and other strategies,the small object and long tail problems in the data set are alleviated.The dataset of road-throwing objects from the top-down perspective has a total of 19,944 images containing road spills and normal road images containing water stains and road markings.The above dataset can provide a specific data basis for the research on the difficulty of object detection from the highway monitoring perspective.(2)Multi-category object detection methods from the monitoring perspective are designed.The multi-scale traffic object detection network from the monitoring perspective is based on the YOLOv4 network.The image road near-far end division module,the small object potential area amplification module and the video frame batch organization module are constructed to solve the problem that it is difficult to continuously detect the object caused by the sharp change of the object shape and scale under the monitoring perspective.This network achieved 84.96%mAP and 58.41 FPS,achieving high-precision and high-efficiency traffic object detection.The road-throwing objects detection network based on anomaly analysis,using the STPM abnormal network,the road image noise reduction module,the global correlation feature calculation module,the sprinkler location extraction module and the automatic training module are constructed,achieving a PRO score of at least 0.924.This network solves the problem of missed detection and wrong detection of various types of road-throwing objects on expressways,and realizes accurate detection of road-throwing objects.(3)A variety of moving object tracking algorithms suitable for single-camera surveillance scenes are proposed.The surveillance view tracking method based on ORB feature points solves the problem that the feature point tracking method is difficult to continuously track the object by extracting the object ORB features,combined with the trajectory distribution strategy,and achieves a MOTA of 49.8%.The multi-scale and multi-feature fusion KCF tracking method combined with DSST,SAMF,and Staple tracking algorithms to extract the features of different dimensions of the object,corrects the defects of the single scale and single feature of the KCF tracker,and achieves a MOTA of 52.6%.The tracking method based on video key frames only uses the object detection results of key frames to calculate the object speed and predict its position,achieving a tracking speed of 41.52 FPS,and reaching 93.07% MOTA in the multiobject tracking dataset of traffic monitoring perspective.Finally,combined with the characteristics of the monitoring scene,the selection methods of the above three tracking methods are designed,and the advantages of each method are fully utilized.In the daytime scene,the three algorithms achieve a MOTA higher than 94%,which provides consistent and correct trajectory results for cross-camera traffic object motion reconstruction.(4)Realized the reconstruction and description of the cross-camera traffic object movement process,generated the global trajectory spatio-temporal map,and analyzed its application in depth.Using the object speed to predict the location of the same object in successive scenes,a coarse temporal calibration method for continuous video streams is designed.Online and offline cross-camera traffic object motion process description methods are designed based on the time coarse calibration results.The online method uses the object’s velocity and spatial position to ensure that the object is matched immediately after it appears in the subsequent scene,achieving at least 85.15% of the adjacent scene trajectory matching rate.The offline method uses the parameters of the fitted object trajectory linear equation,combined with the object appearance similarity measurement results,and has an adjacent scene ID switch rate of up to 8.61%.Both the online and offline global trajectory spatio-temporal map construction methods have a full-scene matching rate of trajectory higher than 80%.Considering the vehicle itself and the surrounding factors,the safety situation of traffic operation is analyzed,and according to the performance of the trajectories in the spatiotemporal map,traffic parameter acquisition methods with a correct rate higher than 90% and traffic incident detection methods with an average traffic incident detection rate of 91.73% are proposed.The reconstruction and application of the cross-camera traffic object movement process can provide real-time primary data for highway safety and smooth traffic flow.According to the above results,it can be seen that this paper can solve the problems of high-precision detection of highway traffic objects and reconstruction of objects motion process under cross-camera,and provide an important reference for the utilization and analysis of highway surveillance videos.