Date of Award:
Doctor of Philosophy (PhD)
Mahdi Nasrullah Al-Ameen
Autonomous vehicles need to gather and understand information from their surroundings to drive safely. Just like how we look around and understand what's happening on the road, these vehicles need to see and make sense of dynamic objects like other cars, pedestrians, and cyclists, and static objects like crosswalks, road barriers, and stop lines.
In this dissertation, we aim to figure out better ways for computers to understand their surroundings in the 3D object detection task and map segmentation task. The 3D object detection task automatically spots objects in 3D (like cars or cyclists) and the map segmentation task automatically divides maps into different sections. To do this, we use attention modules to help the computer focus on important items. We create one network to find 3D objects such as cars on a highway, and one network to divide different parts of a map into different regions. Each of the networks utilizes the attention module and its hierarchical attention module to achieve comparable results with the best methods on challenging benchmarks.
We name the 3D object detection network as Point Cloud Detection Network (PCDet), which utilizes LiDAR sensors to obtain the point cloud inputs with accurate depth information. To solve the problem of lacking multi-scale features and using the high-semantic features ineffectively, the proposed PCDet utilizes Hierarchical Double-branch Spatial Attention (HDSA) to capture high-level and low-level features at the same time. PCDet applies the Double-branch Spatial Attention (DSA) at the early stage and the late stage of the network, which helps to use the high-level features at the beginning of the network and obtain the multiple-scale features. However, HDSA does not consider global relational information. This limitation is solved by Hierarchical Residual Graph Convolutional Attention (HRGCA). PCDet applies the HRGCA module, which contains both graph and coordinate information, to not only effectively acquire the global information but also efficiently estimate contextual relationships of the global information in the 3D point cloud.
We name the map segmentation network as Multi-View Segmentation in Bird's-Eye-View (BEVSeg), which utilizes multiple cameras to obtain multi-view image inputs with plenty of colorful and textured information. The proposed BEVSeg aims to utilize high-level features effectively and solve the common overfitting problems in map segmentation tasks. Specifically, BEVSeg utilizes an Aligned BEV domain data Augmentation (ABA) module to flip, rotate, and scale the BEV feature map and repeat the same process on its ground truths to address overfitting issues. It further incorporates the hierarchical attention mechanisms, namely, HDSA and HRGCA, to effectively capture high-level and low-level features and to estimate global relationships between different regions in both the early stage and the late stage of the network, respectively.
In general, the proposed HDSA is able to capture the high-level features and help utilize the high-level features effectively in both LiDAR-based 3D object detection and multiple camera-based map segmentation tasks, i.e. PCDet and BEVSeg. In addition, we proposed a new effective HRGCA to further capture global relationships between different regions to improve both 3D object detection accuracy and map segmentation performance.
Creative Commons License
This work is licensed under a Creative Commons Attribution-Noncommercial-No Derivative Works 4.0 License.
Chen, Qiuxiao, "Deep Learning With Effective Hierarchical Attention Mechanisms in Perception of Autonomous Vehicles" (2023). All Graduate Theses and Dissertations, Fall 2023 to Present. 58.
Copyright for this work is retained by the student. If you have any questions regarding the inclusion of this work in the Digital Commons, please email us at .