近年来，随着无人驾驶技术的快速发展，行人检测系统已在智慧交通、智能驾驶等领域被广泛关注。在低光照等特殊自然环境条件下，可见光图像由于光照不足、遮挡等因素导致跟踪器性能下降，红外图像由于热辐射信息突出，能够很好的反映出行人目标，但缺乏图像的纹理细节信息。因此，本文提出了基于红外和可见光图像级融合的低光照区域的行人检测方法，对低光照区域的行人检测进一步改善，在保证检测精度的同时，融合图像保留了两幅图像信息以便跟踪器判断和分析，具体工作如下：

（1）针对基于卷积神经网络的图像融合算法依赖卷积核提取图像特征时，感受野会随着卷积层数的增加或卷积深度加深而变小的缺陷，本文提出了基于密集连接和Transformer自编码器的红外和可见光图像级融合方法。Transformer模块从整个图像中提取源图像全局像素之间的交互关系，更多地专注全局特征，密集连接模块更好地了解源图像的局部特征。此外，现有损失函数缺乏图像深层语义信息，因此，网络损失函数引入感知损失项，更好保留了图像中的细节等信息。

（2）针对在低光照区域可见光图像的行人检测出现漏检等问题，红外图像虽然能够提高行人检测的准确度，但是红外图像缺乏周围环境的纹理细节等信息。因此，本文将融合图像作为补充数据集，对全天候场景下低光照等特殊区域的行人进行检测。由于红外目标大多都是远距离拍摄，本文使用CAM模块替代原有YOLOv5的SPPF模块，CAM模块对特征分别以不同的速率进行扩张卷积处理增大感受野，达到提高行人检测精准度的效果。

实验结果表明，本文提出的图像融合方法对比已有的9种图像融合方法，在主观分析和客观指标的评价上都有较优表现。此外，对于在低光照等特殊区域的行人检测，融合图像在检测出行人的同时，也保留了图像红外热辐射和可见光纹理信息。最后，本文所提目标检测方法对行人检测的召回率、准确率、mPA值和F1值相对于其他检测方法都有提升，验证了提出方法的优越性。

In recent years, with the rapid development of driverless technology, pedestrian detection systems have been widely focused on in the fields of intelligent transportation and smart driving. Under special natural environment conditions such as low light, visible images lead to degradation of tracker performance due to insufficient light and occlusion, and infrared images can reflect pedestrian targets well due to prominent thermal radiation information, but lack texture detail information of the images. Therefore, this thesis proposes a pedestrian detection method for low-light areas based on infrared and visible image level fusion to further improve pedestrian detection in low-light areas, and while ensuring detection accuracy, the fused images retain both image information for tracker judgment and analysis, as follows:

(1) To address the drawback that the perceptual field becomes smaller as the number of convolution layers increases or the depth of convolution deepens when the convolutional neural network-based image fusion algorithm relies on convolutional kernels to extract image features, this thesis proposes an infrared and visible image level fusion method based on dense connectivity and Transformer self-encoder. The Transformer module extracts the global pixels of the source image from the whole image between The Transformer module extracts the interrelationship between global pixels of the source image from the whole image and focuses more on the global features, and the Dense Connection module better understands the local features of the source image. In addition, the existing loss function lacks deep semantic information about the image, therefore, the network loss function introduces a perceptual loss term that better preserves information such as details in the image.

(2) To address the problems such as missed detection of pedestrians in visible images in low-light areas, infrared images can improve the accuracy of pedestrian detection, but infrared images lack information such as texture details of the surrounding environment. Therefore, in this thesis, fused images are used as a supplementary dataset for pedestrian detection in special regions such as low light in all-weather scenes. Since most of the IR targets are taken at a long distance, this thesis uses the CAM module to replace the original SPPF module of YOLOv5. The CAM module expands and convolves the features at different rates to increase the perceptual field and achieve the effect of improving the accuracy of pedestrian detection.

The experimental results show that the image fusion method proposed in this thesis has better performance in the evaluation of subjective analysis and objective indexes compared with the existing nine image fusion methods. In addition, for pedestrian detection in special regions such as low-light, the fused images retain the image's infrared thermal radiation and visible texture information while detecting pedestrians. Finally, the recall, accuracy, mPA value, and F1 value of the proposed target detection method for pedestrian detection in this thesis are improved relative to other detection methods, which verifies the superiority of the proposed method.