The Library
Object detection for collision avoidance from lidar point clouds
Tools
Rebane, Martin (2022) Object detection for collision avoidance from lidar point clouds. PhD thesis, University of Warwick.
|
PDF
WRAP_Theses_Rebane_2022.pdf - Submitted Version - Requires a PDF viewer. Download (21Mb) | Preview |
Official URL: http://webcat.warwick.ac.uk/record=b3923738
Abstract
This thesis advances deep learning models that are essential for collision avoidance of autonomous vehicles. The first contribution is an advanced multi-component loss function for 3D object detection algorithms where location of the object and dimensions of its bounding box are estimated simultaneously. The loss function penalises model’s training process when the prediction does not match an expected ground truth. The proposed multi-component loss function enables to observe the progress of locating objects and place greater penalty on bounding box estimation when the object is well located and vice versa. This speeds up the training process as it helps the model to solve the easier task of locating the model first before solving the difficult problem of estimating its bounding box dimensions. Second, a novel sequential point cloud processing method for semantic segmentation is proposed. This uses a sequence of point clouds to generate a prediction. However, as point cloud processing is computationally expensive, processing sequences makes it even more computationally expensive. The proposed method alleviates this problem by fusing point cloud data in a latent feature space instead of processing all point clouds in the sequence each time a new prediction is made. As a result, the method takes advantage of sequential processing while keeping the computational overhead low. Finally, a practical unsupervised method to detect potential collisions in unlabelled point clouds is proposed. The method allows to test the performance and efficiency of different deep learning models on novel data without having to annotate the data first. It is based on the observation that most potential collision areas are defined by the closest object of interest (e.g., a car, a person). Also, the method provides a more realistic assessment of collision probability than widely used aggregate metrics.
Item Type: | Thesis (PhD) | ||||
---|---|---|---|---|---|
Subjects: | Q Science > Q Science (General) Q Science > QA Mathematics > QA76 Electronic computers. Computer science. Computer software T Technology > TK Electrical engineering. Electronics Nuclear engineering T Technology > TL Motor vehicles. Aeronautics. Astronautics |
||||
Library of Congress Subject Headings (LCSH): | Automated vehicles -- Collision avoidance, Automated vehicles -- Computer programs, Optical radar, Deep learning (Machine learning), Sequential processing (Computer science), Image segmentation | ||||
Official Date: | July 2022 | ||||
Dates: |
|
||||
Institution: | University of Warwick | ||||
Theses Department: | School of Engineering | ||||
Thesis Type: | PhD | ||||
Publication Status: | Unpublished | ||||
Supervisor(s)/Advisor: | Tjahjadi, Tardi | ||||
Format of File: | |||||
Extent: | xi, 144 pages : illustrations (some colour) | ||||
Language: | eng |
Request changes or add full text files to a record
Repository staff actions (login required)
View Item |