Work

Vehicle perception with LiDAR and deep learning

Year: 2018
Type: dissertation
Abstract: Autonomous driving is an important area among today's vehicle manufacturers and their suppliers.The development of advanced driver assistance system (ADAS) features that will lead to fully autonomous driving is progressing fast.One of the most important aspects to achieve autonomous driving are through an accurate description of the vehicle's surroundings.As control algorithms, such as trajectory planning, becomes more sophisticated, a comprehensive description of the full surroundings is demanded, including partly and fully occluded objects.In this thesis a solution to accurately predict positions of surrounding objects for highway application using a recurrent neural network is proposed.The system considers raw light detection and ranging (LiDAR) sensor data to compute a final occupancy grid, resulting in an end-to-end solution.Due to the centimeter-level precision of today's LiDARs, we consider during training the LiDAR as ground truth for unoccluded objects and by simulating the input as blocked, and compare the output with a true scan, the network learns to track occluded objects.Hence, no annotated data is required.The LiDAR point clouds are projected into a 2D occupancy grid, which can be seen as an image with one channel, making computer visual techniques such as feature extractions with convolutions possible.The developed networks are built of convolutional and long short-term memory layers and tested on real-world data.We considered five different network structures, where one had the most consistent performance throughout the tests.The obtained result shows that the network is able to accurately predict the location of an object in an occlusion scenario emerged during an overtaking situation.Further, results show that the network performs at approximately 90 % accuracy score after 1.1 s of full occlusion in a highway scenario when the full grid is considered.We conclude that the developed network is able to describe the unoccluded surrounding of the ego vehicle accurately while having the ability to handle shorter times of occlusion.
Language: English
Open Access status: green
OpenAlex ID: W2885228951
Cites: 36
Cited by: 3
Related to: 10