Detection of Moving Objects Based on Efficient Particle Tracking in 1.5D LiDAR Range Images

For autonomous driving it is essential to know the state of dynamic objects in the ego vehicle's vicinity. While learning-based approaches achieve remarkable results in object detection benchmarks, they still struggle to detect unknown object classes, which are not well represented in existing datasets. Occupancy grids in combination with particle tracking have the potential to close this gap for moving objects, as they have only little assumptions about their shape or appearance. Our previous work has shown that particle tracking can be performed very efficiently without occupancy grids but with 1D range images as they implicitly store free space, occupancy and occlusion information. Due to the efficiency of the approach, it is possible to make extensive reasoning about the available free space information, which has been shown to be important in order to distinguish false-positive ghost motion from actual motion. Since the 1D range image only stores the closest obstacle for each azimuth angle of the LiDAR sensor, this approach is not able to detect all moving objects in the sensor's field of view. We propose to use 1.5D range images in combination with an adapted particle tracking to avoid this limitation. This data structure is able to store a variable number of obstacles per azimuth angle. We show that the detection performance is improved as also distant and partially occluded moving objects can be perceived.