Sensor fusion for object detection under adverse weather

Autonomous vehicles (AVs) have rapidly emerged as one of the most rapidly advancing technologies in the realm of transportation. The innovative strides in AV technology, from self-driving cars to autonomous delivery drones, have captured widespread attention and are poised to reshape the future o...

Full description

Saved in:
Bibliographic Details
Main Author: Soh, Brandon Jian Zheng
Other Authors: Soong Boon Hee
Format: Final Year Project
Language:English
Published: Nanyang Technological University 2023
Subjects:
Online Access:https://hdl.handle.net/10356/172746
Tags: Add Tag
No Tags, Be the first to tag this record!
Institution: Nanyang Technological University
Language: English
Description
Summary:Autonomous vehicles (AVs) have rapidly emerged as one of the most rapidly advancing technologies in the realm of transportation. The innovative strides in AV technology, from self-driving cars to autonomous delivery drones, have captured widespread attention and are poised to reshape the future of mobility. Traditional Autonomous Vehicle (AV) camerabased object detectors fail in complicated scenarios where the lighting and weather conditions are not ideal. In recent years, there has been a rise in the use of deep learning methods relying on LiDARs and Radars, given their long history of achieving state of art performance in different types of applications. Despite the rapid development in deep neural networks, object detection is still a persistent challenge as some sensors experience poor perception under adverse weather. Severe weather conditions can still impede LiDAR’s performance despite its excellent capabilities. Snow and water particles make the targets detected by the LiDAR sensor become partially occluded. To address this redundancy, a RADAR sensor is introduced to compensate for LiDAR’s shortcomings. In this project, a sensor fusion approach is proposed to optimize the object detection’s performance in rainy, snowy, or foggy weather. In addition, a comparison of the performance of the object detectors using only one of the proposed sensors will also be performed. Adding on, catastrophic forgetting is a baffling phenomenon that can be observed in neural networks. This happens when a trained neural network has the tendency of losing its ability to make predictions based on its previous training when it is finetuned with new data. Therefore, an investigation will be conducted on the impact of catastrophic forgetting in adverse weather conditions for both LiDAR and RADAR sensors. A comparative assessment will be made to formulate probable resolution.