Learned perception systems for self-driving vehicles
Chaabane, Mohamed, author
Beveridge, Ross J., advisor
O'Hara, Stephen, committee member
Blanchard, Nathaniel, committee member
Anderson, Chuck, committee member
Rebecca, Atadero, committee member
Building self-driving vehicles is one of the most impactful technological challenges of modern artificial intelligence. Self-driving vehicles are widely anticipated to revolutionize the way people and freight move. In this dissertation, we present a collection of work that aims to improve the capability of the perception module, an essential module for safe and reliable autonomous driving. Specifically, it focuses on two perception topics: 1) Geo-localization (mapping) of spatially-compact static objects, and 2) Multi-target object detection and tracking of moving objects in the scene. Accurately estimating the position of static objects, such as traffic lights, from the moving camera of a self-driving car is a challenging problem. In this dissertation, we present a system that improves the localization of static objects by jointly optimizing the components of the system via learning. Our system is comprised of networks that perform: 1) 5DoF object pose estimation from a single image, 2) association of objects between pairs of frames, and 3) multi-object tracking to produce the final geo-localization of the static objects within the scene. We evaluate our approach using a publicly available data set, focusing on traffic lights due to data availability. For each component, we compare against contemporary alternatives and show significantly improved performance. We also show that the end-to-end system performance is further improved via joint training of the constituent models. Next, we propose an efficient joint detection and tracking model named DEFT, or "Detection Embeddings for Tracking." The proposed approach relies on an appearance-based object matching network jointly learned with an underlying object detection network. An LSTM is also added to capture motion constraints. DEFT has comparable accuracy and speed to the top methods on 2D online tracking leaderboards while having significant advantages in robustness when applied to more challenging tracking data. DEFT raises the bar on the nuScenes monocular 3D tracking challenge, more than doubling the performance of the previous top method (3.8x on AMOTA, 2.1x on MOTAR). We analyze the difference in performance between DEFT and the next best-published method on nuScenes and find that DEFT is more robust to occlusions and large inter-frame displacements, making it a superior choice for many use-cases. Third, we present an end-to-end model to solve the tasks of detection, tracking, and sequence modeling from raw sensor data, called Attention-based DEFT. Attention-based DEFT extends the original DEFT by adding an attentional encoder module that uses attention to compute tracklet embedding that 1) jointly reasons about the tracklet dependencies and interaction with other objects present in the scene and 2) captures the context and temporal information of the tracklet's past observations. The experimental results show that Attention-based DEFT performs favorably against or comparable to state-of-the-art trackers. Reasoning about the interactions between the actors in the scene allows Attention-based DEFT to boost the model tracking performance in heavily crowded and complex interactive scenes. We validate the sequence modeling effectiveness of the proposed approach by showing its superiority for velocity estimation task over other baseline methods on both simple and complex scenes. The experiments demonstrate the effectiveness of Attention-based DEFT for capturing spatio-temporal interaction of the crowd for velocity estimation task, which helps it to be more robust to handle complexities in densely crowded scenes. The experimental results show that all the joint models in this dissertation perform better than solving each problem independently.
Includes bibliographical references.
Includes bibliographical references.