面向三维多目标追踪的运动补偿优化方法

Motion Compensation Optimization Method for 3D Multi-Object Tracking

三维多目标追踪是自动驾驶系统中的关键模块之一,其结果的优劣主要取决于追踪模块中数据关联过程的准确度.现有的追踪方法大多从外观特征或运动特征出发计算两帧之间物体的相似度,而基于运动特征的方法通常根据当前帧和历史帧三维包围框之间的交并比(Intersection over Union,IoU)进行关联,然而这种方式在观测点物体自身运动时存在严重缺陷.在观测点物体自身进行运动时,观测到的两帧数据将处于不同的局部坐标系,导致无法使用运动模型准确预测已追踪物体在下一帧中的位置.本文针对上述问题,通过引入观测点自身的惯性测量单元(Inertial Measurement Unit,IMU)或全球定位系统(Global Positioning System,GPS)数据,在一帧数据到达之后计算当前帧局部坐标系与上一帧局部坐标系之间的旋转和平移关系,并对已追踪的物体状态按得到的坐标变换关系进行运动补偿,使其抵消因观测点自身运动造成的偏移量.这种运动补偿增强了追踪模块的数据关联环节,提高追踪时三维包围框的关联成功率,降低误关联数量,改善三维多目标追踪的精度.在相关追踪框架及KITTI数据集上的原型验证表明,所提的运动补偿优化方法实现了1%左右的精度提升.

Three-dimensional(3D)multi-object tracking is a key module in the autonomous driving system,and the quality of the tracking results mainly depends on the accuracy of the data association process in the tracking module.Exist⁃ing tracking methods mostly calculate the similarity of objects between two frames from appearance characteristics or mo⁃tion characteristics,while methods based on motion characteristics usually associate the current frame with the historical frame by using the intersection over union(IoU)of three-dimensional bounding box.However,this method has serious drawbacks when the observation point is moving.When the observation point is moving,the data observed in two frames would lie in different local coordinate systems,making it impossible to use the motion model to accurately predict the posi⁃tion of the tracked objects in the next frame.To solve the above problems,this paper introduces the inertial measurement unit(IMU)or the global positioning system(GPS)data of the observation point itself,and caculates the relationship of rota⁃tion and translation between local coordinate systems of the current and the previous frames after each frame data arrives then the state of the tracked object is compensated according to the obtained coordinate transformation relationship,making it counteract the offset caused by the movement of the observation point itself.This motion compensation enhances the data association process in the tracking module,improving the correlation success rate of the 3D bounding boxes,reducing the number of false correlations,and improving the accuracy of 3D multi-object tracking.The prototype verification on related tracking frameworks and the KITTI dataset shows the proposed motion compensation optimization method achieves an ac⁃curacy improvement of about 1%.