SPAD-Based LiDAR Data Pre-Processing on Embedded Systems Using Machine Learning

With the rise of Advanced Driver Assistance Systems (ADAS), range sensors and corresponding data processing methods are becoming more and more important. It plays a decisive role in one of the major future topics, autonomous driving. Light Detection And Ranging (LiDAR) sensors are attracting attention due to their unique advantages in terms of high radial distance resolution. Therein, Single Photon Avalanche Diode (SPAD)-based direct Time-Of-Flight (TOF) LiDAR system shows a powerful ability regarding its simple measurement principle and extremely high energy detection efficiency. This system measures the distances based on the traveling time of laser photons to reach and return to an object. However, one of the greatest challenges of such a system is high background light, which causes a number of false triggers and interferes with the desired signals. Moreover, due to the large volume of LiDAR data, classical methods are only able to perform simple processing on low-level LiDAR data in applications with a high demand on timing performance, such as ADAS. In the subsequent high-level processing, only the depth information (i.e. point cloud) is utilized and the used LiDAR front-end is considered as a black box. In this case, most features in low-level LiDAR data are overlooked, resulting in low system robustness in harsh conditions, such as high background light and large distances.
This work seeks a breakthrough on the distance determination performance of a SPAD-based direct TOF LiDAR system. In the scope of the work, a machine learning-based distance prediction approach, Multi-Peak Analysis (MPA), is proposed for LiDAR time-correlated histograms to improve timing performance and the system robustness against background light. First of all, the existing methods based on the LiDAR system are reviewed. Concepts and attempts are designed and carried out on the combination of machine learning and low-level LiDAR data. Two datasets are used to assess the performance of MPA. After theoretical and simulative evaluations, MPA is designed and includes three components: 1) a physics-guided feature extraction, 2) two distance prediction algorithms (fully-connected neural network and naive Bayes classifier), and 3) a correlation analysis in time and space. The evaluation results lead to the conclusion that MPA outperforms a widely-used classical digital processing in terms of timing performance and distance determination with a +-5% error bound with various background light intensities and distances. Finally, MPA is implemented on both a personal computer and a Field-Programmable Gate Array (FPGA) module for runtime verifications. Demonstrations show a high agreement with theoretical and simulative results and confirm the feasibility of MPA.



Citation style:
Could not load citation form.