The multi object detection method of fused images based on improved YOLOv7

DOI:	10.11809/bqzbgcxb2023.06.023
Keywords:	object detection; YOLOv7; BoT; image fusion; deep learning
Abstract:	Aiming at the low accuracy for object detection in a low light environment, this paper proposes a multi object detection method based on improved YOLOv7 of visible low light and infrared fused image. Combining the advantages of visible light and infrared images, this paper makes a fused image dataset by adopting Fusion by a Generative Adversarial Network (FusionGAN). It also introduces Bottleneck Transformer (BoT) to YOLOv7 model to make the network pay more attention to the overall image information, and improves the feature extraction capability so as to increase the accuracy of pedestrian and car detection. Besides, it improves the regression loss function from CIoU to SIoU, which reduces the degree of freedom and accelerates convergence of the network. Then, an improved YOLOv7 algorithm is obtained, namely, BoT YOLOv7. Experiments are conducted on public datasets of Visible Infrared Paired dataset for Low Light vision (LLVIP）and Multi Spectral Road Scenarios (MSRS), and the results show that BoT YOLOv7 has higher detection accuracy for fused images than for visible light or infrared images. The improved algorithm achieves an average accuracy of 92.6% for the fusion images, which is 5.83% higher than that of the original YOLOv7 model. Both missing and false detection rates are low for detecting targets like pedestrians and cars through BoT YOLOv7 algorithm, which indicates that the improved algorithm has good accuracy and real time performance, and can meet the requirement of multi object detection in a low light environment.
Published:	2023-06-28
PDF HTML