The closest section to the vehicle is the car tip margin, which is the distance from the center of the robot to the point of lateral sampling, the length of which determines the smoothness of steering when switching between trajectories.
The next section is called the roll-in margin, which is the distance from the outer limit of the car tip margin to the point of parallel lateral sampling, the length of which is proportional to the vehicle’s velocity. The faster the vehicle is traveling, the longer this section should be to generate smooth change.
The section farthest from the vehicle is called the roll-out section, which runs from the outer limit of the roll-in zone to the end of the length of the local trajectory.
原始参数设置是在:op_trajectory_generator这个节点
有待看具体的代码 看看单位是m嘛?如果是m会不会太长了一点
下面这里最大规划距离maximum planning distance应该怎么确定?
在common参数里有Path distance,确定一下是否是这个
这里的垂直路径点是指?垂直于什么的?垂直区域的话就直接指向天了
这里指出了横向距离是0,所以为啥还要写垂直路径点,直接longitudinal不就行了?
使用局部路径规划算法生出roll-outs的时候有三个步骤
利用现在的位置和最大规划距离从全局路径中提取interest区域
sample 新的垂直路径点对应于全局路径中提取的区域,sample的点是从car-tip margin处出发的,但是横向的 lateral距离是0,然后再sample延伸到roll-in区域的边界,再根据有的路径计算roll_out density
Transition cost: constrains the vehicle from jumping roll-outs 这可以使得转弯更顺滑,计算方法就是normalized perpendicular distance between roll-outs and currently selected trajectory
关于输入的障碍物,Autoware的实现在core_perception包里,两种类型:bounding boxes 和cluster of point cloud
Bounding boxes的一直没有输出 建议再次检查一下原因 April 28, 2021
直接拿了groud_truth来做的; July 3, 2021 后面做纯点云识别的时候,发现可能是点云的稀疏程度的原因,也就是Autoware本身拿的64线激光雷达,但是我们是16线的(Carla仿真当时用的也是16线的)
为了保证精度和performance
bounding boxes 可以提高检测障碍物的过程
Cluster 可以提高精度但是degrades performance drastically
所以trade-off是:使用小部分sample的contour point from cluster,然后每个障碍物最多用16个点来表示【咦咦咦咦!!】
题外话:如果是传统机器人方法是可以直接cluster of point cloud然后走到costmap再到global 是混合A*
这个max number of contour point也是可以调整的一个参数
Max number of contour point: increasing this number we can achieve finer representation, lead more accurate obstacle avoidance