Deep learning for object detection under rainy conditions

• Main Author: Chin, Zhuo Sheng
• Other Authors: Teoh Eam Khwang
• Format: Final Year Project
• Language: English
• Published: 2018
• Subjects: DRNTU::Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision; DRNTU::Engineering::Computer science and engineering::Computing methodologies::Artificial intelligence
• Online Access: http://hdl.handle.net/10356/74597
• Institution: Nanyang Technological University

Fast, robust and accurate object detection are required for autonomous driving. While the main technology used for obstacle detection and avoidance is RADAR and LIDAR, LIDAR is affected by rain and snow, while RADAR resolution is low due to its longer wavelength. Vision based object detection is widely adopted because it is cost-effective and have a wide field-of-view. Unfortunately, current object detectors are not designed for adverse weather conditions. The performance of existing vision-based object detection methods, e.g. Mobileye, drops significantly under rainy conditions. This motivates us to develop an object detector which is robust to heavy rain. Furthermore, object detection developed in this report can be used to supplement RADAR and improve accuracy. In this project, the author aims to develop an object detection system which can provide high accuracy while under rainy conditions. This will be done by collecting a dataset with vehicles in rain scenes, and training a state-of-the-art deep learning model using this dataset. By adding these training data, fast, robust and accurate stereo object detection can be attained. Our results have showed that training deep learning models on rain scenes does indeed improve their accuracy in rainy situations, faring much better than simple rain removal via image sharpening. This technique is generalisable for training any deep learning model.