一种新的结合卷积神经网络的隧道内停车检测方法

为了更准确地检测高速公路隧道内停车行为,将传统的图像处理技术与深度学习相结合.首先,通过基于混合高斯模型(Gaussian mixture model,GMM)的背景差分法提取出运动目标.接着通过meanshift算法跟踪这些运动目标,计算运动目标的速度以及在相邻视频帧中运动目标的相关性,通过与速度阈值和相似度阈值的比较得到静止目标.最后,结合卷积神经网络(convolutional neural network,CNN)分类模型识别静止目标是否为车辆.文中方法在高速公路隧道视频上进行测试,达到至少84％的准确率.另外,与没有结合卷积神经网络的传统图像处理方法相比,文中方法至少提高了63％的准确率.

A new tunnel vehicle stopping detection methodology combined with convolutional neural network

In order to more accurately detect the vehicle stopping in highway tunnels, this paper proposes a new methodology that combines the traditional image processing technology with deep learning. Firstly, the foreground moving targets are extracted using the background difference method based on Gaussian mixture model (GMM). Then the meanshift algorithm is applied to track these foreground moving targets. By calculating the speed of the moving targets and the correlation of the moving targets between the neighboring video frames, and comparing the results with the speed threshold and correlation threshold, the static target is detected. Finally, combined with the convolutional neural network (CNN) classification model, whether the static target is vehicle is identified. The method proposed in this paper is validated using the real highway tunnel vehicle stopping video and achieves an accuracy of at least 84%. Compared with the traditional image processing method without CNN, our method improves at least 63% accuracy.