煤矸识别是煤矿智能化的重要方向之一，有效的识别有助于提高煤矿安全性，避免资源浪费提高经济效益。近年来，我国正积极推动煤炭行业绿色发展，为了满足绿色环保的要求，迫切需要更加智能、环保和高效的煤矸分选技术。与传统识别方法相比，深度学习技术能够保证煤矸分选工作高效、节能并避免环境污染的发生。因此，本文基于YOLOv5对煤矸目标检测及跟踪方法进行研究，具体研究内容如下：

(1)提出一种基于YOLOv5的目标检测方法。在空间金字塔中引入空洞卷积，不损失图像信息的前提下增大卷积输出感受野，强化深层特征的提取；添加卷积注意力模块，快捷高效地分析复杂场景信息；采用AdaBelief优化器提高模型的收敛速度与识别精度。实验结果表明：改进后模型mAP值达到94.43%，比原YOLOv5模型提高了2.27%，在极端黑暗环境中也能准确划定目标边界。

(2)采用YOLOv5+DeepSort的目标跟踪方法。利用带加速度的卡尔曼滤波器对目标状态进行预测和更新，更好地处理目标的加速度和运动变化，提高跟踪准确性。通过数据关联匹配算法将检测框和跟踪框进行匹配，提高跟踪的一致性。跟踪实验结果表明，该方法相对于DeepSort模型在MOTA和MOTP指标上分别提高了3.6%和4.4%，能够更好地处理被遮挡和小目标煤矸，提高跟踪性能。

(3)通过模型剪枝和知识蒸馏对算法进行压缩优化。稀疏化训练引入稀疏约束，筛选出对网络性能影响较小的通道，减少模型冗余和复杂度。通过混合通道剪枝进一步对网络进行剪枝，得到轻量化的网络作为学生网络。最后通过知识蒸馏的方式，将改进后网络作为教师网络，向学生网络传递知识。通过知识蒸馏保证模型精度，提高模型的部署能力。实验结果表明当稀疏率为0.005，剪枝率为0.5时，混合剪枝后的模型性能最佳，模型大小降为剪枝前的50%。仅为13.1MB，平均推理速度达到34FPS，mAP值达到90%以上，跟踪中没有发生失败情况，验证了网络轻量化的有效性。

Effective identification of coal gangue can improve coal mine safety, avoid resource waste and improve enterprises economic benefits.It is necessary to put forward more intelligent and efficient separation technology of coal gangue.This paper will study the detection and tracking method of coal gangue target based on YOLOv5. Specific research contents are as follows:

(1)Dilated convolution is introduced into space pyramid,convolutional attention module is added to analyze complex scene information. AdaBelief optimizer is used to improve the convergence speed and recognition accuracy of the model. The experimental results show that the mAP of the improved model reaches 94.43%, 2.27% higher than that of the original YOLOv5 model.

(2)The combined model of improved YOLOv5+DeepSort is adopted in this paper. The improved model is used to detect the target, the Kalman filter with acceleration predict the tracking frame, and the data association matching algorithm is used to correlate the detection frame with the tracking frame. The experiment results show that compared with the DeepSort model, the MOTA and MOTP of the improved model are increased by 3.6% and 4.4%.

(3)Proposed model compression.Sparse training results in structured sparseness of channels, and channels that have little impact on network performance are screened out.A global pruning strategy is adopted. The normally trained associative network is used as the teacher network, and the pruned model is used as the student network for knowledge distillation. Model compression improves the detection speed while ensuring accuracy.