Lidar sensors are a core element of fully automated vehicles. The reliable detection of surrounding objects depends very much on the algorithm used. The classic approach uses segmentation based on a distance criterion. This methodology is very robust, but also has disadvantages with low point cloud densities and for fusion with other sensors. Alternatives are grid-based algorithms or machine learning methods. Here too there is a dependency on the sensor measurement principle. Therefore, a comparative study was carried out on selected scenarios using all three methods for three different lidar sensors. To calculate the accuracy of the predictions, different metrics have been used such as OSPA Metric for tracks, IOU, Mean Average Precision, and Average Orientation Similarity for bounding boxes accuracy. This makes an objective comparison of the detection results possible. The robustness of the algorithms and the essential parameters is also discussed.
Session:
Poster
|