For tracking, autonomous vehicles often use 3D laser range finders (LRFs), which are expensive. In order to make autonomous cars affordable as a mass product, we use multiple, more affordable 2D LRFs for tracking and implement an algorithm with the goal of achieving a similar performance. Our tracking algorithm comprises the following steps: Data preprocessing, segmentation, classification, feature extraction, data association, and track management. To compensate the information reduction, we re-use track information and focus on reducing over-segmentation and effects of the shape change problem. We test the algorithm on own labeled data and training data sets of “The KITTI Vision Benchmark Suite” with the CLEARMOT and MT/PT/ML metrics. In addition, we test specific properties with specific scenarios. Our algorithm obtains a MOT accuracy, which reflects the amount of correctly detected objects, of 0.79 on our own and negative values on the KITTI data sets. Obtained MOT precision, which is an averaged detection precision, is around 60 percent on all data sets. Out of all tracks, 47.83 percent are mostly tracked on our own data and zero percent on KITTI data sets. Our algorithm for tracking multiple objects with multiple moving 2D LRFs does not reach 3D performance. Improvements of feature extraction and especially classification to distinguish between vehicle and non-vehicle objects would boost its performance.
«
For tracking, autonomous vehicles often use 3D laser range finders (LRFs), which are expensive. In order to make autonomous cars affordable as a mass product, we use multiple, more affordable 2D LRFs for tracking and implement an algorithm with the goal of achieving a similar performance. Our tracking algorithm comprises the following steps: Data preprocessing, segmentation, classification, feature extraction, data association, and track management. To compensate the information reduction, we re...
»