Extraction automatique par apprentissage profond des obstacles et des facilitateurs à la mobilité des personnes à mobilité réduite à partir des données LiDAR mobile
|Advisor:||Mostafavi, Mir Abolfazl; Guilbert, Éric|
|Abstract:||Mobility is a fundamental life habit for the social participation of people with motor disabilities (PMD). One of the biggest challenges for PMDs is to find accessible itineraries for their movement in the city. In this respect, several research groups, including MobiliSIG, are interested in assessing the accessibility of places to support the development of mobility assistance tools for PMDs. However, traditional methods for acquiring and processing data relevant to the analysis of the accessibility of the urban environments are generally inefficient and very costly in terms of time and money. In this context, the lidar technology presents an interesting alternative for the acquisition of very detailed and accurate data on the urban environment. Moreover, artificial intelligence techniques have shown great potential for the automatic extraction of relevant information from lidar point clouds. To this end, the overall objective of this research is to evaluate the potential of new deep learning-based approaches for the semantic segmentation of lidar point clouds to automate the extraction of obstacles and facilitators (sidewalks, island, steps, etc.) related to the mobility of PMDs. To do so, we were particularly interested in the potential of deep learning methods such as Superpoint graph and FKAconv algorithms. The main steps of this research are: 1) to develop an annotated 3D database dedicated to mobility setoff PMDs, 2) to apply and evaluate the deep learning algorithms, 3) to highlight the challenges encountered in 3D semantic learning (irregular and voluminous data, complexity of urban scenes, highly variable morphology of instances, etc.). The selected algorithms are applied to mobile lidar data to analyze access to shops in downtown Quebec City. The results of this research have demonstrated the potential of deep learning methods for semantic segmentation of elements relevant to PRM mobility from mobile lidar data. However, these methods still suffer from several problems that lead to misclassifications leading to segmentation imperfections.|
|Document Type:||Mémoire de maîtrise|
|Open Access Date:||13 December 2021|
|Collection:||Thèses et mémoires|
All documents in CorpusUL are protected by Copyright Act of Canada.