In recent years, LiDAR has emerged as a crucial perception tool for robotics and autonomous vehicles. However, most LiDAR perception methods are adapted from 2D image-based deep learning methods, which are not well-suited to the unique geometric structure of LiDAR point cloud data. This domain gap poses challenges for the fast-growing LiDAR perception tasks. This dissertation aims to investigate suitable deep network structures tailored for LiDAR point cloud data, and therefore design a more efficient and robust LiDAR perception framework. Our approach to address this challenge is twofold. First, we recognize that LiDAR point cloud data is characterized by an imbalanced and sparse distribution in the 3D space, which is not effectively captured by traditional voxel-based convolution methods that treat the 3D map uniformly. To address this issue, we aim to develop a more efficient feature extraction method by either counteracting the imbalanced feature distribution or incorporating global contextual information using a transformer decoder. Second, besides the gap between the 2D and 3D domains, we acknowledge that different LiDAR perception tasks have unique requirements and therefore require separate network designs, resulting in significant network redundancy. To address this, we aim to improve the efficiency of the network design by developing a unified multi-task network that shares the feature-extracting stage and performs different tasks using specific heads. More importantly, we aim to enhance the accuracy of different tasks by leveraging the multi-task learning framework to enable mutual improvements. We propose different models based on these motivations and evaluate them on several large-scale LiDAR point cloud perception datasets, achieving state-of-the-art performance. Lastly, we summarize the key findings of this dissertation and propose future research directions.

Environment perception is crucial to autonomous driving. Panoptic segmentation and objects tracking are two challenging tasks, and the combination of both, namely 4D panoptic segmentation draws researchers’ attention recently. In this work, we implement 4D panoptic LiDAR segmentation (4D-PLS) on Volvo datasets and provide a pipeline of data preparation, model building and model optimization. The main contributions of this work include: (1) building the Volvo datasets; (2) adopting an 4D-PLS model improved by Hyperparameter Optimization (HPO). We annotate point cloud data collected from Volvo CE, and take a supervised learning approach by employing a Deep Neural Network (DNN) to extract features from point cloud data. On the basis of the 4D-PLS model, we employ Bayesian Optimization to find the best hyperparameters for our data, and improve the model performance within a small training budget. / Miljöuppfattning är avgörande för autonom körning. Panoptisk segmentering och objektspårning är två utmanande uppgifter, och kombinationen av båda, nämligen 4D panoptisk segmentering, har nyligen uppmärksammat forskarna. I detta arbete implementerar vi 4D-PLS på Volvos datauppsättningar och tillhandahåller en pipeline av dataförberedelse, modellbyggande och modelloptimering. De huvudsakliga bidragen från detta arbete inkluderar: (1) bygga upp Volvos datauppsättningar; (2) anta en 4D-PLS-modell förbättrad av HPO. Vi kommenterar punktmolndata som samlats in från Volvo CE och använder ett övervakat lärande genom att använda en DNN för att extrahera funktioner från punktmolnsdata. På basis av 4D-PLS-modellen använder vi Bayesian Optimization för att hitta de bästa hyperparametrarna för vår data och förbättra modellens prestanda inom en liten utbildningsbudget.