Automotive Target Models for Point Cloud Sensors

Abstract

One of the major challenges to enable automated driving is the perception of other road users in the host vehicle’s vicinity. Various automotive sensors that provide detailed information about other traffic participants have been developed to handle this challenge. Of particular interest for this work are Light Detection and Ranging (LIDAR) and Radio Detection and Ranging (RADAR) sensors, which generate multiple, spatially distributed, noise corrupted point measurements on other traffic participants. Based on these point measurements, the traffic participant’s kinematic and shape parameters have to be estimated. The choice of a suitable extent model is paramount to accurately track a target’s position, orientation and other parameters. How well a model performs typically depends on the type of target that has to be tracked, e.g. pedestrians, bikes or cars, as well as the sensor’s setup and measurement principle itself. This work considers the creation of extended object models and corresponding inference strategies for tracking automotive vehicles based on accumulated point cloud data. We gain insights into the extended object model’s requirements by analysing automotive LIDAR and RADAR sensor data. This analysis aids in the identification of relevant features from the measurement’s spatial distribution and their incorporation into an accurate target model. The analysis lays the foundation for our main contributions. We developed a constrained Spline-based geometric representation and a corresponding inference strategy for the contour of cars in LIDAR data. We further developed a heuristic to account for the integration of the measurement distribution on cars, generated by LIDAR sensors mounted on the roof of the recording vessel. Last, we developed an extended target model for cars based on automotive RADAR sensors. The model provides an interpretation of a learned Gaussian Mixture Model (GMM) as scatter sources and uses the Probabilistic Multi-Hypothesis Tracker (PMHT) to formulate a closed form Maximum a Posteriori (MAP) update. All developed approaches are evaluated on real world data sets.