| Unmanned vehicles help to reduce traffic congestion in urban areas and improve traffic efficiency and accessibility to public transportation services.Unmanned vehicles use autonomous environment sensing technologies to detect and track potentially dangerous objects and trajectory planning and control technologies to achieve autonomous navigation.Therefore,research on environment perception systems with reliable target detection and tracking abilities is important to improve the safety of unmanned vehicles.Current research on target detection and tracking based on vision and LiDAR has achieved some results.However,in real traffic environments,constrained by complex and variable road traffic conditions and unpredictable weather conditions,current target detection and multi-target tracking algorithms still face issues such as reduced detection rates,tracking accuracy,precision,and real-time performance.To address the above problems,this dissertation focuses on target detection and tracking in traffic scenes and,based on existing research,carries out research on target detection algorithms,motion state estimation algorithms with adaptive time-varying noise and multi-target tracking(MTT)algorithms in traffic scenes to provide a proven solution for accurate and reliable environment sensing of unmanned vehicles in traffic scenes.The main work and innovations of the dissertation are as follows:(1)Improved two-step fusion target detection method with lateral inhibition mechanism-based object proposal generation.Addressing the task of target detection in traffic scenarios characterized by numerous target types,reduced LiDAR localization accuracy,and significant variations in ambient lighting,an improved two-step fusion detection methodology integrated with vision-based object proposal generation is proposed.The method uses the matching results of LiDAR detection and visual object proposal as the ROI and uses the R-CNN model to classify the ROI,thus reducing the impact of LiDAR localization accuracy on visual recognition performance.To address the problem that current visual object proposal generation methods have difficulty achieving a good balance between recall,localization quality,robustness,and computational efficiency,a proposal generation method OPM(Object Proposal and Merge)based on the human eye lateral inhibition mechanism is proposed.This method uses constructed enhanced frequency features to describe objects,an improved binarization strategy to accelerate the computation of the classifier,a cascade classification strategy combining features and dimensions to improve the classification accuracy,and a merging strategy based on the lateral inhibition mechanism to improve the quality of the final generated object proposal box.The OPM method improves the performance balance problem of current object proposal generation algorithms,reduces the visual detection performance in two-step fusion over-dependence on LiDAR point cloud detection accuracy,and improves the real-time,detection precision of object detection.On the VOC2007 dataset,the average processing time of the OPM method for each image frame is only 0.0014 s,and the average robustness can reach 80%,the recall rate can reach 99.3%,and the localization quality can reach 81.1%.Utilizing the refined two-step fusion method,the LiDAR point cloud detection outcomes of Voxel R-CNN on the KITTI dataset witnessed a 6.95%enhancement in detection accuracy.Experimental validation conducted with the autonomously developed "Intelligent Pioneer" vehicle in urban road traffic contexts yielded detection recall and accuracy rates of 94.7%and 99.6%,respectively.(2)Design of adaptive time-varying noise filter and motion state estimation.A new NC2(Normalized estimation,Calibration,and Correction)filtering algorithm with adaptive time-varying noise is proposed to address the problem of degraded tracking performance due to time-varying measurement noise and process noise of the tracking filter in traffic scenes.This algorithm estimates the covariance matrices of the process and measurement noises by analyzing the residual sequence of the previous N steps and uses the constructed self-covariance calibrator and Gaussian calibrator to jointly diagnose the noise bias,thereby correcting the noise bias through a negative feedback mechanism.The filtering algorithm can simultaneously estimate the covariance matrices of the process and measurement noises in the tracking system and the target motion state,improving the issues with current adaptive filtering algorithms such as narrow application scope,stringent technical assumptions,inadequate real-time performance,and poor robustness.Experimental results show that the NC2 algorithm has significant advantages in estimating the covariance of process and measurement noises,with its average estimation error significantly reduced compared to other methods,demonstrating its excellent performance in precision and stability.(3)Multi-target tracking(MTT)based on the NC2 adaptive filter and 3D cascaded association method.Investigated the 3D MTT algorithm in road traffic scenes and its key techniques,namely target association and motion state estimation methods.Focusing on the characteristics of a large number of target detection results,dense targets,and high real-time requirements of MTT systems in traffic scenes,a 3D cascaded target association method is proposed.By integrating this method with the two-step fusion method and the NC2 adaptive filter,a new 3D multi-target tracking algorithm,FNC2(Fusion and NC2),is proposed,effectively enhancing the accuracy and real-time performance of multi-target tracking in complex traffic environments.Tests on the KITTI multi-target tracking dataset show that the FNC2 algorithm has high operational efficiency,reaching 91.6fps,and in the public evaluation on the KITTI official server in June 2021,FNC2’s overall performance ranked second on the KITTI pedestrian multi-target tracking leaderboard,surpassing other 3D multi-target tracking methods on the leaderboard.Furthermore,the algorithm is applied to the "Intelligent Pioneer" unmanned vehicles,conducting experimental validation in various traffic scenarios such as urban,suburban,and wilderness,with multi-target tracking accuracies reaching 94.6%,96.1%,and 99.7%respectively,proving its superior performance and application potential.(4)Harsh environment target detection and tracking based on multisensor fusion and adaptive tracker.To address the problem that vision sensors are easily obscured by smoke,a robust target detection method based on vision and millimeter wave radar sensor fusion is proposed,which uses the property that millimeter wave radar can penetrate smoke to compensate for the lack of vision sensors obscured by smoke,and then uses the fused data from vision and millimeter wave radar for target tracking to compensate for visual missed detection.Experimental results in smoke scenarios show that the method has an average pixel error of 12.1 pixels and a robustness of 88.1%,outperforming the compared methods.Additionally,to address the issue of LiDAR being affected by rain,causing omissions and changes in measurement noise,a sequential fusion strategy of LiDAR and millimeter-wave radar is proposed to improve the reliability of target detection.The NC2 filter is used to handle noise variations,further enhancing the tracker’s performance.Experiments conducted in rainy conditions validated the effectiveness of the method,achieving 0.29m in positioning accuracy and 0.21m/s in speed estimation accuracy,demonstrating the method’s strong adaptability and high accuracy under adverse weather conditions. |
PDF Full Text Request