AI-Based Point Cloud Upsampling for Autonomous Driving Systems

Autonomous driving, decades ago relegated to the realm of science fiction, emerged as a tangible reality that is rapidly transforming the automotive industry, redefining our relationship with vehicles, and placing them in the spotlight of both the industry and the general public. Through the study and analysis of modern and efficient interpolation techniques, we aim to reduce the current costs and processing requirements associated with the LiDAR sensor, which is one of the main information sources. Our approach explores the fusion of lower-cost LiDAR sensors with advanced interpolation techniques, with a particular focus on achieving performance parity with pricier 64-channel LiDAR setups. This work is based on 3 main axes: firstly, the analysis of available LiDAR data and its representation; secondly, the development and implementation of an interpolation technique based on 1D convolutional layers integrated with fully connected layers, in order to analyse data coming from a sliding window; and finally, the comparative evaluation of the results between different state-of-the-art interpolation techniques, using object detection networks in point clouds. By interpolating the point clouds with the proposed technique, improvements between 1.92% and 30.98% in detection and classification tasks were achieved, depending on the object and the type of detection (3D or bird's eye view). Furthermore, computational efficiency was not left aside by reducing the inference times necessary for interpolation, compared to other techniques used as contrast. This highlights the viability and scalability of our approach in realizing cost-effective yet high-performance autonomous driving systems.