摘要
为提高无人车行驶过程中前方车辆检测的准确性和实时性,提出了基于激光雷达(LIght Detection And Ranging,LIDAR)深度信息和视觉方向梯度直方图(Histograms of Oriented Gradients,HOG)特征的车辆识别和跟踪方法。目标首次进入视野时,聚类处理激光雷达深度信息并确定假设目标的候选区域,采用车辆尾部的HOG特征对假设目标进行验证。在HOG特征验证前,基于最小二乘支持向量机(Least Squares Support Vector Machine,LS-SVM)算法对样本集HOG特征进行训练学习,生成车辆分类器模型。对于验证后的目标车辆,采用激光雷达获取的深度信息对目标车辆进行持续跟踪。构建了2种车辆模型,结合最小二乘直线拟合方法提取出车辆特征,生成目标模型。同时,提出了基于多特征马氏距离的目标关联代价方程,实现了多目标的关联;完成了基于卡尔曼滤波的车辆状态滤波和位置估计,更新了跟踪器模型。通过有效的管理策略,实现了目标跟踪的3个状态:1)初始化模型的生成;2)跟踪过程中跟踪器的更新与预测;3)目标驶离视野时跟踪器的删除。最后,通过试验验证了跟踪算法的有效性。
In order to improve the veracity and instantaneity of preceding vehicle detection in the running process of UGV( Unmanned Ground Vehicle),a vehicle detection and tracking method based on depth information and visual HOG( Histograms of Oriented Gradients) feature of LIDAR( LIght Detection And Ranging) is proposed. When the target vehicle enters the vision of testing field for the first time,the LIDAR depth information is processed by clustering and the candidate area of the hypothetical target is determined,and the hypothetical target is verified by the HOG feature of the vehicle's tail. Before the HOG feature verification,the HOG feature of the sample set is trained and learned based on the Least Squares Support Vector Machine( LS-SVM) algorithm to generate the vehicle classifier model. For the verified target vehicle,the depth information acquired by the LIDAR is used to continuously track the target vehicle. Two vehicle models are constructed,and the vehicle features are extracted by the least square line fitting method to generate the target model. In addition,the target correlation cost equation based on multi-feature Mahalanobis distance is proposed to achieve multi-target correlation. The Kalman filtering is used to estimate the vehicle filter and position,and the tracker model is updated. Three statesof target tracking are realized by effective management strategy: generation of target initialization tracker,tracker update and prediction,and removal of the tracker when the target is out of view. Finally,the effectiveness of the tracking algorithm is verified by experiments.
出处
《装甲兵工程学院学报》
2017年第6期89-96,共8页
Journal of Academy of Armored Force Engineering
