Perception using multi-tasked neural networks on ATLASCAR2

• Autor(a) principal: Ribeiro, Gonçalo Manuel Cordeiro
• Data de Publicação: 2023
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10773/40926
• topic: Deep learning; Artificial intelligence; Multitask; Multiple models; Object detection; Image segmentation; ATLASCAR2

Efficient perception is a fundamental requirement for ADAS and ADS, with implications for safety, accuracy, and speed. The choice between multitasked and single-tasked deep learning networks can significantly impact the performance of these systems and their ability to understand and respond to the complex driving environment. This dissertation explores the comparison between multi-tasked neural networks and multiple single-tasked networks. It investigates car perception, focusing on object detection and image segmentation, covering car detection, road segmentation, and lane marking. To make the comparisons possible and also to implement different kinds of models in the ATLASCAR2’s inference unit, a versatile software system designed to seamlessly run multiple deep-learning models with distinct tasks was developed for this dissertation. Single-tasked networks like YOLOv5, YOLOv7, and YOLOv8 were evaluated for object detection, while road segmentation was evaluated with Mask2Former, UPerNet, and SegFormer. Lane marking was analyzed using RESA, O2SFormer, and UFLDv2. The multi-tasked networks evaluated included YOLOP, YOLOPv2, and TiwnLiteNet. The dissertation findings indicate that combining multiple single-tasked models can lead to synchronization challenges and slower inference speeds. Multi-tasked networks outperform multiple single-tasked models in terms of efficiency, although their performance benefits are more pronounced when handling tasks that share a closer relationship.