The ever-improving technology in the field of automotive LiDAR has introduced a range of newer and more advanced sensor technologies that are significantly more capable than their predecessors. Manual labeling of objects in the point cloud and drawing bounding boxes is a highly time-consuming, resource-intensive, and expensive task. With the unveiling of a newer version of the LiDAR sensor every few years, this task becomes redundant and should be eliminated. By transforming the data recorded from previous-generation sensors into that of the next generation, we can avoid the laborious process of manually labeling data, the high cost of driving thousands of kilometers, and save valuable time.This research work leverages various deep learning techniques to enhance the quality and resolution of low-density point clouds, enabling more detailed reconstructions of the complex objects using range images. In particular, the model utilizes a multi-scale convolution neural network(CNN) architecture to extract feature representations from the input point clouds, followed by a series of up-sampling layers that generate high-resolution point clouds with improved spatial coherence and structural fidelity.This research significantly contributes to the advancement of point cloud processing techniques, enhancing their efficiency and accuracy. These advancements hold great potential for diverse fields, including autonomous driving and robotics.
Session:
Autonomous Driving
|
| 14:10 - 14:40