The paper addresses the problem of detecting pedestrians using three dimensional data acquired by an autonomous mobile robot equipped with an on-board 3D laser scanner. Previous works in this field have dealt with various approaches for combining 2D and 3D range data features for the use in pedestrian classification. In this paper we propose an image processing pipeline for generating a depth image from point clouds data and then localizing object candidates from the depth image. It involves the image segmentation, feature extraction and human classification processes within unstructured dynamic environments. Three different approaches for the detection of pedestrians, vehicles and cyclists using only 3D range data were employed as a part of this system. We train and test the classifiers in an open environment, with presence of multiple pedestrians, cyclists and vehicles, using only point cloud data. The effectiveness and robustness of the proposed system are verified through experiments with real data. This system is also capable to deal with a real-time framerate (10Hz) with high accuracy.

One of the major application areas of highly automated vehicles is the problem of Automated Valet Parking (AVP). In this work, we analyze solutions and compare performances of RRT (rapidly exploring random tree) based approaches in the context of the AVP problem, which can also be applied in a more general low-speed autonomy context. We present comparison results using both simulation and real-life experiments on a representative parking use case. The results indicate better suitability of RRTx and RRV for utilization in typical AVP scenarios. The main contributions of this work lie in real-life experimental validation and comparisons of RRT approaches for use in low-speed autonomy.

A reliable and precise model of the environment is of the highest importance for autonomous vehicles. Occupancy grids are a well-known approach for environment modeling and are a crucial part of multiple autonomous driving functionalities. The standard method is to use a single 2D occupancy grid to model the environment using nonground points. In this paper, we propose a decentralized occupancy grid filtering chain (pipeline) where a high-density 64-layer LiDAR provided the input to our pipeline. Our approach enables us to obtain detailed 2D and 3D models of the environment simultaneously. The pipeline was validated on different scenarios in both simulation and real world. The performance of the designed occupancy grid pipeline was evaluated by the proposed key performance indicators (KPIs) based on accuracy. The results have shown that the approach was able to extract free space information with a high degree of accuracy, while reducing the size of the unobserved area in the grid compared to the standard methods and achieving real-time performance.