目前许多矿山对于矿卡司机的不安全行为监督仍依赖于人为监管，无法及时准确地发现问题，利用计算机技术识别不安全行为是替代人工检测的一条高效途径。本文利用深度学习来解决视频序列的矿卡司机不安全行为识别，深度学习方法不依赖人工设计特征，而是自适应地学习更好的高维特征，具有稳健性更好、速度更快及准确率更高的优点。首先，对帧图像采用翻转、旋转和添加噪点等方法进行数据增强，以降低样本的不均衡性；其次，利用本文优化的模型训练数据。结果表明：网络测试准确率达到93.445%，相比原始双流网络模型提高了15%。将本文模型与不考虑时序动态信息的深度学习模型进行试验比较，证明了时域特征信息对于行为识别的重要性。综上，本文提出的网络模型对于矿卡司机不安全行为的识别率较高，对矿卡司机不安全行为的识别及采矿生产作业安全具有重要实践意义。

At present，many mines still rely on human supervision to supervise the unsafe behavior of mining truck drivers，and cannot find problems timely and accurately.This consumes a certain amount of manpower and material resources but cannot solve the problem.With the development of computer technology and artificial intelligence technology，more and more fields are beginning to use artificial intelligence technology to supervise the unsafe behavior of mining truck drivers，such as intelligent security，unmanned driving，and intelligent transportation.Behavior recognition is a hot issue in the field of computer vision.Using computer technology to identify unsafe behaviors is an efficient way to replace manual detection.This paper uses deep learning to solve the unsafe behavior recognition of mining truck drivers in video sequences.The traditional deep learning method does not rely on artificial design features，but adaptively learns better high-dimensional features，better robustness，and faster speed，the accuracy rate is higher.Firstly，according to the actual obtained video data，by analyzing the relative position relationship between the camera and the driver’s area，the video is clipped to obtain video data with less redundant information.At the same time，in order to reduce the imbalance of the data samples，by using flipping，methods such as panning and adding noise were used to enhance the data set，and then use Opencv to re-convert the enhanced image data into a video file and use the dense_flow method to obtain an optical flow diagram.Secondly，use the network for training and testing.In order to conduct com-parative experiments，firstly，a traditional classification model that does not consider time sequence information was used for training and testing，and the transfer learning method was used to train Resnet，Xception，and Inception.And fusion of three single models to get a new fusion model.At the same time，the time domain and spatial domain channels of the dual-stream network model are set to the pre-trained VGG16 using migration learning under the consideration of timing information，and the comparison experiment was carried out with the C3D-two-stream proposed in this paper.The experimental results show that the improved Vgg-two-stream model can reach an accuracy rate of 89.539%，and the accuracy rate of the C3D-two-stream model can reach 93.445%.In summary，the C3D-two-stream model proposed in this paper has a high recognition rate.It also proves that for behavior recognition，the acquisition of characteristic information in the time dimension can make the recognition results more accurate，which has important practical significance for the recognition of unsafe behaviors of mining truck drivers.

http://www.goldsci.ac.cn/article/2021/1005-2518/1005-2518-2021-29-1-14.shtml