Design and Evaluation of Data Dissemination Algorithms to Improve Object Detection in Autonomous Driving Networks

In the last few years, the amount of information that is produced by an autonomous vehicle is increasing proportionally with the number and resolution of sensors that cars are equipped with. Cars can be provided with cameras and Light Detection and Ranging (LiDAR) sensors, respectively needed to obtain a two-dimensional (2D) and three-dimensional (3D) representation of the environment. Due to the huge amount of data that multiple self-driving vehicles can push over a communication network, how these data are selected, stored, and sent is crucial. Various techniques have been developed to manage vehicular data; for example, compression can be used to alleviate the burden of data transmission over bandwidth-constrained channels and facilitate real-time communications. However, aggressive levels of compression may corrupt automotive data, and prevent proper detection of critical road objects in the scene. Along these lines, in this thesis, we studied the trade-off between compression efficiency and accuracy. To do so, we considered synthetic automotive data generated from the SELMA dataset. Then, we compared the performance of several state-of-the-art algorithms, based on machine learning, for compressing and detecting objects on LiDAR point clouds. We were able to reduce the point cloud by tens to hundreds times without any significant loss in the final detection accuracy. In a second phase, we focused our attention on the optimization of the number and type of sensors that are more meaningful to object detection operations. Notably, we tested our dataset on a sensor fusion algorithm that can combine both 2D and 3D data to have a better understanding of the environment. The results show that, although sensor fusion always achieves more accurate detections, using 3D inputs only can obtain similar results for large objects while mitigating the burden on the channel.