Multimodal perception for autonomous driving

Autonomous driving is set to play an important role among intelligent transportation systems in the coming decades. The advantages of its large-scale implementation –reduced accidents, shorter commuting times, or higher fuel efficiency– have made its development a priority for academia and industry. However, there is still a long way to go to achieve full self-driving vehicles, capable of dealing with any scenario without human intervention. To this end, advances in control, navigation and, especially, environment perception technologies are yet required. In particular, the detection of other road users that may interfere with the vehicle’s trajectory is a key element, since it allows to model the current traffic situation and, thus, to make decisions accordingly. The objective of this thesis is to provide solutions to some of the main challenges of on-board perception systems, such as extrinsic calibration of sensors, object detection, and deployment on real platforms. First, a calibration method for obtaining the relative transformation between pairs of sensors is introduced, eliminating the complex manual adjustment of these parameters. The algorithm makes use of an original calibration pattern and supports LiDARs, and monocular and stereo cameras. Second, different deep learning models for 3D object detection using LiDAR data in its bird’s eye view projection are presented. Through a novel encoding, the use of architectures tailored to image detection is proposed to process the 3D information of point clouds in real time. Furthermore, the effectiveness of using this projection together with image features is analyzed. Finally, a method to mitigate the accuracy drop of LiDARbased detection networks when deployed in ad-hoc configurations is introduced. For this purpose, the simulation of virtual signals mimicking the specifications of the desired real device is used to generate new annotated datasets that can be used to train the models. The performance of the proposed methods is evaluated against other existing alternatives using reference benchmarks in the field of computer vision (KITTI and nuScenes) and through experiments in open traffic with an automated vehicle. The results obtained demonstrate the relevance of the presented work and its suitability for commercial use.