Multispectral sample augmentation and illumination guidance for RGB-T object detection by mm detection framework

Abstract

Multispectral object detection technology has important application prospects in the fields of autonomous driving and so on. Conventional multispectral object detection algorithm rely solely on deep neural networks to learn multispectral image sample information, lacking the guidance of prior knowledge, and not fully utilizing infrared, visible, and other spectral information, resulting in decreased accuracy of object detection in complex scenes. To address this problem, this paper proposes an object detection algorithm based on infrared visible sample augmentation and illumination guidance. The algorithm adopts the MMDetection framework and extracts multispectral object features based on a designed sample augmentation method based on the fusion of positive and negative samples in multispectral images. Based on a designed adaptive weight allocation method guided by illumination, it enhances the algorithm's adaptability to the lighting environment. Finally, through the design of a multi-task loss function, it achieves high-precision and robust object detection in complex scenes. Experimental results on datasets such as FLIR and M3FD show that the proposed algorithm has significant advantages over comparative algorithms such as CFR_3 and GAFF in terms of average detection precision.