车辆检测是计算机视觉领域备受瞩目的技术之一，它可以应用于无人驾驶、交通流量信息统计、智能停车等领域，从而有助于有效缓解交通问题，提升民众生活质量。然而现实生活中的交通路段车辆密集，车辆目标经常被大面积遮挡，导致漏检和误检的情况频发。同时相机在拍摄车辆照片时，光线和天气条件往往成为影响检测准确性的关键因素。针对上述问题，本文以YOLOv7算法为基础，设计构建了一种高效且准确的道路车辆目标检测方法，主要研究内容如下：

（1）为解决车辆遮挡和远距离小目标车辆检测精度不佳的问题，引入多卷积融合机制，即将深度可分离卷积（Depthwise Separable Convolution）和空洞卷积（Dilated Convolution）融入金字塔池化模块（Pyramid Pooling Module, PPM）中，构建出YOLOv7-SDC模型。该结构可以整合多个不同尺度的特征图，弥补因环境复杂和目标尺度不一而丢失的车辆特征信息，同时降低计算复杂度和增大感受野，从而提高车辆检测精度。

（2）为了提高车辆检测的准确性并减少外界光线的干扰，本文引入了注意力机制，使模型能够突出车辆重点特征，抑制次要特征，提高网络的特征提取能力。具体做法是，将混合注意力机制（Convolutional Block Attention Module, CBAM）的串行连接形式，设计为并行连接结构，并借鉴高效通道注意力机制（Efficient Channel Attention Network, ECANet）的思想，结合残差网络，设计出一种适用于光线较差场景下的车辆检测模型YOLOv7-CASM，从而使模型更轻量化、稳定和高效。

（3）为验证模型的有效性，对不同场景（车辆密集与稀疏、光线充足与较差）和不同类别（小、中、大目标）进行了测试对比。实验结果显示，本文方法在车辆密集环境、光线较差条件和远距离小目标场景下的精度提升明显。为验证模型的泛化能力，基于三个不同数据集进行实验，与SSD、DETR、Deformable DETR、YOLOv7以及YOLOv8相比，本文方法不仅在检测精度上表现出色，而且在检测速度方面取得了良好的平衡。总体而言，该方法在精度和速度之间取得了优异的表现。

综上所述，本文的YOLOv7-SDC-CASM模型能够满足多种场景下的车辆检测任务，可为后续研究提供数据支持，具有较高的实用价值。

Vehicle detection is one of the high-profile technologies in the field of computer vision, which can be applied to unmanned driving, traffic flow information statistics, intelligent parking and other fields, thus helping to effectively alleviate traffic problems and improve the quality of life of the people. However, real-life traffic sections are densely populated with vehicles, and the vehicle target is often obscured by a large area, leading to frequent leakage and misdetection. Meanwhile, when the camera is taking photos of vehicles, light and weather conditions often become the key factors affecting the accuracy of detection. Aiming at the above problems, this paper designs and constructs an efficient and accurate road vehicle target detection method based on YOLOv7 algorithm, and the main research contents are as follows:

(1) In order to solve the problems of vehicle occlusion and poor vehicle detection accuracy for small targets at long distances, a multi-convolution fusion mechanism is introduced, i.e., Depthwise Separable Convolution and Dilated Convolution are integrated into the Pyramid Pooling Module (PPM) to construct the YOLOv7-SDC model. This structure can integrate multiple feature maps of different scales to make up for the vehicle feature information lost due to the complex environment and different target scales, while reducing the computational complexity and increasing the receptive field, thus improving the vehicle detection accuracy.

(2) In order to improve the accuracy of vehicle detection and reduce the interference of external light, this paper introduces the attention mechanism, so that the model can highlight the key features of the vehicle, suppress the secondary features, and improve the feature extraction capability of the network. Specifically, the serial connection form of Convolutional Block Attention Module (CBAM) is designed as a parallel connection structure, and the idea of Efficient Channel Attention Network (ECANet) is borrowed. Combined with the residual structure, a vehicle detection model YOLOv7-CASM for poorly lit scenarios is designed, thus making the model more lightweight, stable, and efficient.

(3) In order to verify the effectiveness of the model, test comparisons are conducted for different scenarios (dense vs. sparse vehicles, well-lit vs. poorly lit) and different categories (small, medium, and large targets). The experimental results show that the accuracy of this paper's method is significantly improved in dense vehicle environments, poorly lit conditions, and long-distance small target scenarios. To verify the generalization ability of the model, experiments are conducted based on three different datasets, and the method in this paper not only performs well in terms of detection accuracy, but also achieves a good balance in terms of detection speed when compared with SSD, DETR, Deformable DETR, YOLOv7, and YOLOv8. Overall, the method achieves an excellent performance between accuracy and speed.

In summary, the YOLOv7-SDC-CASM model in this paper is able to satisfy the vehicle detection tasks in a variety of scenarios, which can provide data support for the subsequent research and has high practical value.