Convolutional Sparse Coding and Dictionary Learning for Lidar Depth Completion in Automotive Scenarios

The interest in LiDAR sensors for autonomous driving applications has recently increased: solid state architectures have made it possible to reduce sizes and costs while maintaining higher scanning resolution compared to RADAR sensors. In a dynamic automotive scenario, LiDAR depth measurements result in a discrete point cloud which is typically sparse and may contain irregular/missing depth information, moreover, one may further reduce the illuminated pixels to increase the scanning rate and reduce computational cost. Here, a novel application of both a patch based and a convolutional sparse coding (CSC) approach for LiDAR depth completion are presented and validated using the publicly available KITTI dataset of outdoor automotive scenarios. Patch-based sparse coding approach may result inaccurate in representing global image features and edges, especially when the missing data percentage is high. CSC allows to process the data globally while still preserving local information by constructing the dictio-nary as a concatenation of convolutional filters. The dictionary considered for both approaches is either composed of Daubechies Wavelets or learned from depth images of the urban SYNTHIA dataset using K-SVD and Convolutional Dictionary Learning (CDL) strategies. Resulting depth maps using the CSC based approach for various sparsity levels produce smooth images and an enhanced scenario awareness. An analysis based on the Sparse Mean Absolute Error (SMAE) and Weighted Mean Absolute Error (WMAE) indicates that depth and edge preservation improves with respect to patch-based strategies.