Deep learning-based object detection for autonomous vehicle under rainy conditions

• Main Author: Teh, Kelvin Kae Wen
• Other Authors: Soong Boon Hee
• Format: Final Year Project
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Electrical and electronic engineering::Computer hardware, software and systems
• Online Access: https://hdl.handle.net/10356/149860
• Institution: Nanyang Technological University

An autonomous vehicle (AV) is a vehicle that have the ability to manoeuvring safely outside controlled environment with no human interventions in controlling the vehicle. AV uses multiple types of sensors, most commonly Light Detection and Ranging (LIDAR) and video camera to understand the environment around the vehicle. As video camera are relatively cheaper as compared to LIDAR, it is adopted in most vehicles with semi-autonomous driving capabilities today. In order for a vehicle to achieve a high level of autonomy, vehicle is required to operate safely under a whole slew of road and weather conditions regardless of any adverse conditions. Visual feed of autonomous vehicles might be obscured or tainted during rainy conditions when raindrops that landed on these camera fails to be cleared away when the vehicle is in motion or stationery. As it is not always possible to remove these raindrops stains from the visual feed, this project aims to address the problem by creating a more robust object detection framework that is able to detect other road vehicles under rainy conditions. This will be accomplished by implementing a key-point detection algorithm to detect road objects together with an object detection model such as Faster Region-Convolutional Network (Faster R-CNN) and YOLOv3. The novel key-point detection algorithm will be used to detect tail lights of vehicles within the field of view. The features are then enhanced and passed to an object detection model to perform the object detection task. This allows for the object detection framework to better detect other road vehicles under rainy conditions. The implementations of tail light detection algorithm with the object detection model shows an improvement over the baseline model performance and achieve 40.3% accuracy on the challenging rain dataset.