Accepted by DSN 2024:

Demo

  • Enhanced the resiliency of AVs in unfamiliar and accident-prone scenarios with a novel traffic risk assessment method
  • Validated the method by unit-testing a prototype using designed experiments in CARLA Simulator and Argoverse dataset, including real-world geometric and semantic metadata, lane boundaries, geometric LiDAR, and ring camera information
  • Created multi-threaded data generation and testing pipelines and boosted efficiency by 200% using subprocess in Python
  • Engineered a memory-efficient ResNet variant, reducing the footprint by 50% while achieving 95%+ testing accuracy
  • Constructed 6000 scenarios from NHTSA pre-crash typologies and trained lightweight Double DQN Reinforcement Learning Agents in PyTorch to preemptively brake using the traffic risk as an indicator, reducing accidents by 72.7%
  • Cleansed and statistically analyzed generated datasets using pandas, and visualized outcomes with seaborn and Matplotlib