Hall of Mirrors
Adaptive Segmentation of LiDAR Data Enabling Efficient Object Detection and Localization (304947)
Monica Martinez-Canales, Intel Corp.*Rita Chattopadhyay, Intel Corp.
Keywords: LiDAR, Autonomous vehicles, Adaptive, Object Detection and Localization, Adaptive Segmentation
Light Detection And Ranging (LiDAR) is increasingly being used for object detection by autonomous vehicles. An essential step in LiDAR data processing is; Segmentation. Segmentation methods identify clusters corresponding to different objects (cars, cyclists, pedestrians, poles, trees etc.) in data. However the greatest challenge in segmenting LiDAR data, is the uneven distribution of the LiDAR returns in the 3D space. We observe that as distance increases, the number of LiDAR returns decrease significantly. This change in LiDAR data distribution poses a great challenge to any clustering or segmentation algorithm. This is because segmentation algorithms require user inputs defining a cluster; minimum number of points (k) within a neighborhood and radius of the neighborhood (Eps). However, since the LiDAR data distribution changes with range, hence using fixed k or Eps values do not provide satisfactory results at all ranges. The paper proposes a solution to the problem, by computing both k and Eps parameters adaptively for each range, at real time, thus enabling efficient object detection at all ranges. Secondly, the method automatically computes k and Eps, making it scalable.