智能化是煤矿安全生产和快速发展的必然趋势，构建智能化生产系统的框架、云边计算协同体系，针对煤矿重点目标进行快速、精准的目标检测，是煤矿智能化发展的重要方向，然而现有的基于深度学习的目标检测算法体积庞大、计算复杂度高，难以将其部署到边缘设备来提供低时延数据分析服务。提出一种基于YOLO-v4L-EA算法的煤矿重点目标检测方法，将融合空间注意力机制的轻量级YOLO-v4部署到矿山物联网系统的边缘计算设备，从而为用户提供高响应速度的重点目标感知服务。算法层面，针对煤矿井下图像存在亮度不均匀等缺陷影响目标检测效果的问题，设计像素正则化空间注意力结构(PNSAM)，用批量正则化实现空间注意力机制，该结构能够辅助目标检测模型加强对重要特征关注，有助于算法从低质量图像中感知任务目标；受MobileNet基础结构启发，基于深度可分离卷积对YOLO-v4骨干网络执行轻量化改进，使整体模型能够部署在矿井边缘计算设备；为减小h-swish激活函数造成的梯度损失，在模型中尝试使用Mish激活函数，凭借其梯度光滑特性来实现高效的深层特征提取。根据煤矿视频监控数据构建矿山目标检测数据集，用于评价目标检测网络模型实际应用性能。使用NVIDIA Jetson TX2边缘计算平台作为试验硬件设备，对比试验显示该模型在公测数据集VOC2012的试验结果相对于YOLO-v4-Tiny模型，mAP值提升了13.39%，证明该算法的有效性和正确性；该模型在矿山目标检测数据集上的mAP值为88.9%，说明该方法能有效实现煤矿井下重点目标检测。

Intelligence is an inevitable trend for the safe production and rapid development of coal mines, and it is an important direction for the development of coal mine intelligence to build the framework of intelligent production system, cloud-edge computing synergy system, and fast and accurate object detection for key targets in coal mines. However, the existing deep learning-based object detection algorithms are large in size and high in computational complexity, making it difficult to deploy them to edge devices to provide low-latency data analysis services. The article proposes a coal mine focused object detection method based on the YOLO-v4L-EA algorithm, which deploys the lightweight YOLO-v4 incorporating the spatial attention mechanism to the edge computing devices of the mine IoT system, so as to provide users with focused object sensing services with high response speed. At the algorithmic level, to address the problem of uneven brightness and other defects in coal mine underground images that affect the effect of object detection, the article designs the pixel regularized spatial attention structure (PNSAM), which implements the spatial attention mechanism with batch regularization, which can assist the target detection model to strengthen the attention to important features and help the algorithm to perceive the task target from low-quality images; inspired by the MobileNet base inspired by the MobileNet base structure, the YOLO-v4 backbone network is lightweight and improved based on the depth-separable convolution, so that the overall model can be deployed in the mine edge computing devices; in order to reduce the gradient loss caused by the h-swish activation function, this paper tries to use the Mish activation function in the model, which achieves efficient deep feature extraction by virtue of its gradient-smoothing property. A mine object detection dataset is constructed based on coal mine video surveillance data for evaluating the practical application performance of the object detection network model. The article uses the NVIDIA Jetson TX2 edge computing platform as the experimental hardware equipment, and the comparison experiments show that the experimental results of the model on the public test dataset VOC2012 improve the mAP value by 13.39% relative to the YOLO-v4-Tiny model, which proves the validity and correctness of the algorithm; the mAP value of the model on the mine object detection dataset is 88.9%, indicating that the method can effectively realize the key object detection in coal mine underground.