"Visual Motion of Vehicle Interactions for Autonomous Driving"

"Vehicles move on roads by performing regular actions on the ground, taking into account the road environment and traffic rules. Vehicle cameras capture such motion of surrounding vehicles on the ego-vehicle in a video. According to the visual appearance of vehicles and their motion trajectories, human drivers can identify the interactions of surrounding vehicles during stable vehicle action\/driving without difficulty. This dataset collects visual data of driving scenes in videos, and we label the qualitative vehicle actions around the camera based on the detailed motion trajectories of surrounding vehicles. The data are used in the deep learning of temporal events and vehicle behaviors in the sensing of autonomous vehicles. Vehicle interactions are limited to a range of 50m. Depending on different lanes and relative speed around the ego-vehicle, 13 classes of qualitative motion (interaction) are specified. They are identified in the video for preparing ego-vehicle action and avoiding collision. This is particularly important for fast-moving vehicles to plan vehicle motion and path ahead, and respond promptly in accident avoidance. To obtain the motion of vehicles moving on a horizontal plane, the video is sampled at the height of the horizon to produce the Motion Profile at 30fps temporally, which contains the trajectories of vehicles and background scenes. Along with the Motion Profile, the width of the vehicle in the images, which reflects the depth of the vehicle, is also obtained and plotted with the Motion Profile, as input to a spatial-temporal deep network to understand vehicle interactions. The motion trajectories are annotated with the interaction classes in specified colors, separated at the moment when a new interaction starts."

车辆依托道路环境与交通规则，在路面执行规范动作以完成道路行驶。车载摄像头可在视频中采集自车（ego-vehicle）视角下周边车辆的运动画面。人类驾驶员可借助车辆视觉外观与运动轨迹，在稳定行驶过程中轻松识别周边车辆的交互行为。本数据集采集了驾驶场景的视频视觉数据，并基于周边车辆的精细运动轨迹，对相机视野内的车辆动作进行定性标注。该数据集可用于自动驾驶感知场景下的时序事件与车辆行为深度学习任务。车辆交互范围限定在50米以内。结合自车周边的车道条件与相对速度情况，本数据集共定义了13类定性运动（交互）类别。在视频中识别这些类别，可辅助自车规划行驶动作并规避碰撞。这对于高速行驶车辆提前规划运动与行驶路径、及时响应以避免事故尤为关键。为获取车辆在水平面的运动信息，本数据集以地平线高度为基准对视频进行采样，生成帧率为30fps的时序Motion Profile，其中包含车辆与背景场景的轨迹信息。同时，我们还提取了图像中反映车辆深度的车身宽度信息，并将其与Motion Profile结合，作为时空深度学习网络的输入，以实现车辆交互行为的理解。运动轨迹将以指定颜色标注对应的交互类别，并在新交互开始的时刻进行分段标注。