Multi-Scale Deep Learning Frameworks for Trajectory Prediction, Scene Understanding, and Perception in Autonomous Driving
Year: 2026
Type: dissertation
Abstract: Reliable autonomous vehicle operation at urban intersections requires competence across the prediction–perception stack, yet current deep learning models underuse the structural priors that govern real-world driving. This thesis argues that incorporating structural priors at multiple scales improves the accuracy, interpretability, and robustness of autonomous driving models, through five interconnected studies. Chapter 3 introduces a Turn-Aware LSTM that encodes turning intent via cumulative heading change. On UAV footage from a signalised intersection in Châteauguay, Québec, it reduces Final Displacement Error by 15–20% for turning manoeuvres at a 3-second horizon. Chapter 4 establishes lane graph conditioning as an architecture-flexible inductive bias for trajectory prediction. A lightweight module (under 50,000 parameters), evaluated on 89,258 Waymo scenarios, reduces ADE by 9.3% at 3 s and, at 8 s, minADE by 26.6%, minFDE by 32.6%, and miss rate by 42.7%. Chapter 5 recovers real-world locations for anonymised scenarios by matching intersection topologies to OpenStreetMap via WayGraph's 48-dimensional star pattern descriptor. Across 56,797 scenarios it achieves 90% top-1 accuracy under synthetic OSM-to-OSM matching (33–55% under realistic noise), and a geographic audit reveals fourfold under-sampling of signalised intersections (4.3% vs. 17.8% city-wide). Chapter 6 develops a spatial attention visualisation framework that projects Transformer attention onto bird's-eye-view scenes. Applied to MTR-Lite, it identifies a tunnel vision pattern: failed predictions are associated with lower attention entropy (5.72 vs. 5.94 bits) and elevated self-attention; cyclist targets show higher miss rates (88.1% vs. 54.0%). Left-turn manoeuvres trigger a general agent-to-map attention shift (−19.9% vehicle attention, p < 0.001), suggesting possible VRU-related attention vulnerabilities, though the cyclist-specific interaction did not reach significance (p = 0.51, n = 17) and requires validation on larger samples. Chapter 7 shows that viewpoint diversity suppresses false positives via dual-camera LiDAR fusion with vehicle- and drone-mounted cameras. Across 2,600 CARLA frames and ten seeds, symmetric late fusion improves mAP by +0.92 percentage points (p = 0.001) and reduces false positives by 13% on a representative seed. These studies support the thesis argument across five scales—kinematic, road topology, geographic, attentional, and sensor geometry—and outline future research directions.
Author: Zhou, Xingnan
Institution: Concordia University
Language: English
Subfield: Automotive Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
Funder: Mitacs
OpenAlex ID: W7221486811