Loading...

Scene Interaction Aware Perception for Autonomous Driving Tasks

Farokhnejad Afshar, Mehrnaz | 2023

0 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: English
  3. Document No: 57084 (52)
  4. University: Sharif University of Technology, International Campus, Kish Island
  5. Department: Science and Engineering
  6. Advisor(s): Ghafourian Ghahramani, Amir Ali; Shirmohammadi, Zahra
  7. Abstract:
  8. Safely navigating complex urban environments presents a critical challenge in autonomous driving perception tasks. This challenge necessitates the ability to consider interaction with traffic scene agents while accurately predicting the behavior of vulnerable road users (VRUs) like pedestrians. This thesis aims to enhance AV perception by focusing on scene interaction awareness from two key perspectives: vehicle-vehicle and vehicle-pedestrian interactions. For vehicle-vehicle interactions, monocular depth estimation, a low-cost, data-driven approach, is employed to approximate inter-vehicle distance from an RGB image. First, vehicles and their lights are detected using the YOLOv7 algorithm then these detected information are mapped to radial depth using a nonlinear function. The attention mechanism is also included to improve detection accuracy. Results on the KITTI dataset (mean relative distance error of 17%) indicate that the approach is competitive and superior to current state-of-the-art models. In addressing vehicle-pedestrian interactions, this thesis proposes an efficient model for predicting pedestrian crossing intentions. This model extracts 2D keypoints utilizing the Open Pif-Paf network and uses a skeleton-based transformer to anticipate pedestrian actions. By focusing on 2D pose features and attention mechanisms, the model avoids the complexities of 3D representations while remaining adaptable to diverse lighting conditions. Results show this model achieves higher precision (0.86), F1 (0.87), and accuracy (0.82) than existing models and offers a more efficient and robust solution for understanding pedestrian-vehicle interactions in urban areas
  9. Keywords:
  10. Autonomous Vehicles (AVs) ; Object Detection ; Depth Estimation ; Perception ; Human Pose Estimation ; You Only Look Once (YOLOv7) ; Pedestrian Intention Prediction ; Human Behaviour Analysis

 Digital Object List

 Bookmark

No TOC