Segmentation, Classification and Tracking of Objects in LiDAR Point Cloud Data Using Deep Learning

Sammanfattning: The purpose of this thesis was to explore deep learning methods of segmentation, classification and tracking of objects in LiDAR data. To do this a complete pipeline was developed, consisting of background filtering, clustering, tracking, labeling and visualization. The objects that were focused on were pedestrians, cyclists, cars and animals, in different environments. Background segmentation and object detection was done using classical methods, using distance filtering and DBSCAN for clustering. Four deep neural networks were trained for object classification and two for semantic segmentation, with different parameters to compare performance. To process point clouds generated by a LiDAR, a specialized architecture was needed, which is why PointNet layers were used to build the models. Tracking was managed by a recurrent neural network structure, capable of predicting object trajectory, updating measurements and data association. An additional classical tracking algorithm, was also developed as a baseline for comparison. The networks were trained purely on simulated data from an autonomous driving simulator, with the aim of also functioning on real world data. The models were compared and evaluated on both simulated data and real world LiDAR data. The results showed that classification using PointNet as foundation works well, even on real world data, being able to accurately classify both humans and vehicles. Semantic segmentation proved not to be suitable for the task, lacking in performance. The deep learning tracker showed great potential, but was difficult to properly train to outperform the classical tracker.