| Resumo: | Road accidents are one of the leading causes of death, with drivers being responsible for 90 percent of these. The most viable solution to save lives is to move on to autonomous driving, which explains why such a technology has been intensively investigated. An autonomous vehicle must first be aware of its surroundings. At present, such is done entirely using vision sensors, which capture information both from the scene – the background – and from dynamic objects – the foreground. This thesis explores a different approach in which the background needs not be observed, as the vehicle already carries detailed maps describing the scene. Vision sensors thus have the sole purpose of reconning dynamic objects. Given that such an approach enables discarding background information from sensor data, only foreground data that really matters is processed, leading to higher precision in detecting objects as well as to increased computational efficiency. The vision sensor that best suits the proposed approach is the LiDAR. First, unlike camera images, the point clouds generated by a LiDAR are not projections. As a result, a point that exists both in the generated point cloud and in the detailed maps belongs to the background and may thus be discarded. Second, a point cloud has more dimensions than the image captured by a camera, making it harder to process. It is therefore important to reduce the number of points of points clouds before processing these. This thesis provides the first-ever demonstration of the proposed approach. For the sake of completeness, such a demonstration is done resorting to two datasets: a dataset comprising point clouds captured by a real LiDAR – a real dataset – and a dataset comprising synthetically generated point clouds – a synthetic dataset. All results confirm that removing background points from a point cloud decreases the computational effort involved in 3D object identification while increasing its average precision.
|
|---|