Enhancing Object Detection Methods by Knowledge Distillation for Automotive Driving in Real-World Settings
Date of Award
Ph.D. in Electrical Engineering
Department of Electrical and Computer Engineering
Commercial cameras primarily aim to capture visually appealing images for human viewers, often leading to the loss of critical information during the image generation process. However, for machine vision applications, extracting as much data as possible from an image is crucial for effective operation. In the context of autonomous vehicles, cameras serve as vital vision tools, where data captured is processed through object detection algorithms such as YOLO, FasterRCNN, RetinaNet, etc. Hence, it becomes essential to have an object detection algorithm capable of leveraging all available information from camera images to perform effectively under challenging conditions, such as low-light scenarios and the detection of small or distant objects. Traditionally, the establishment and evaluation of most object detection models have been based on common RGB images, which align with human visual perception. However, important details that could be valuable for machine vision tasks often vanish through the image signal processing (ISP) pipeline. To address this limitation, cameras with an RCCB (Red, Clear, Clear, Blue) color format, replacing the green channel with clear, have been introduced in the autonomous driving industry featuring more low-light sensitivity and less noise absorptive; which leads to enhanced object detection quality. This research focuses on training cost-effective object detection models 3 using raw images captured with an RCCB color filter array, while requiring a minimum amount of training data and low computational complexity. The author employs a knowledge distillation method through unsupervised learning to transfer the knowledge from high-performance state-of-the-art object detection models, trained on RGGB (Red, Green, Green, Blue) color filter array images, to operate with high accuracy on RCCB raw images. The results of this study demonstrate the effectiveness of the proposed approach in training object detection models specifically tailored for autonomous driving applications. By leveraging RCCB raw images and incorporating knowledge distillation, a compelling performance was achieved while optimizing training costs and computational requirements.
Automotive Driving- Knowledge Distillation - Object Detection - Raw Image - RCCB Color Filter Array
Copyright © 2023, Author
Kian, Setareh, "Enhancing Object Detection Methods by Knowledge Distillation for Automotive Driving in Real-World Settings" (2023). Graduate Theses and Dissertations. 7274.