Evaluation of data augmentation techniques in 3D object detection for self-driving vehicles

• Autor(a) principal: Santos, Xavier Pinto
• Data de Publicação: 2021
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10773/34129
• topic: Autonomous driving; Data augmentation; Computer vision; LiDAR sensors; Point cloud; 3D object detection deep learning; Machine learning; KITTI; PointPillars

Traffic accidents are one of most common causes of unnatural death in the world. Emerging new technologies have been improving the safety of the vehicles and reducing the mortality rate on crashes. The automation of the driving task will eventually remove human error from the equation and reduce this number even further. Automating the task of driving requires that vehicles observe their surroundings. Such a task is done by vision sensors, of which LiDAR plays a key role. LiDARs produce high-definition point clouds, which are subsequently used for detecting objects. To this function is given the name of 3D Object Detection in the field of Computer Vision. Most Object Detection models nowadays are based on Deep Learning algorithms, which require massive amounts of data to be trained. The retrieval of this data is not a trivial task. Cars must be equipped with an array of sensors and driven around different scenarios for hours on end. The retrieved data then needs to be painstakingly annotated by humans. Being such an expensive and time demanding process to gather a dataset, one should make sure to fully harness the information contained in each sample. Such can be done through Data Augmentation. The process of Data Augmentation contributes significantly for the improvement of the performance gains, being actually more relevant than the advances in the object detection models themselves. This work reports the first-ever detailed quantification of the impact that the inclusion of augmented samples in a dataset has in 3D Object Detection in the context of autonomous driving. Only well established tools were used in order to enable a direct comparison with state of the art results. Such tools include the KITTI dataset, the PointPillars 3D Object Detection model, the metric of Average Precision, and proven Data Augmentation strategies. The obtained results confirm that suitably augmenting data significantly improves the Average Precision in detecting objects in 3D.