基于改进CNN的公交车内拥挤状态识别

Recognition of Crowded State in Bus Based on Improved Convolution Neural Network

针对传统的视频图像处理方法对公交车内乘客拥挤状态的检测受运动阴影、动态背景及场景光照变化等因素的影响问题,提出了一种基于改进卷积神经网络VGG-16的公交车内拥挤状态识别方法。该方法在VGG-16的模型基础上,优化全连接层层数,使用迁移学习共享VGG-16预训练模型的各层权值参数进行训练。相对于文中的传统图像处理方法、AlexNet模型、GooleNet模型以及标准VGG-16模型,改进的VGG-16模型对公交车拥挤状态的识别准确率最高,识别精度能够达到96.1%。模型的损失值比标准VGG-16模型收敛得更快,模型表现得更加稳定。实验证明:改进后的VGG-16模型能够更好地提取公交内拥挤状态的特征,解决公交车内拥挤状态的识别问题。

Aiming at the problem that the traditional video image processing method is used to detect the crowded state of passengers in the bus,such as motion shadow,dynamic background and scene illumination changes,a crowded bus state recognition method based on improved convolution neural network VGG-16 is proposed. Based on the VGG-16 model,this method optimizes the number of all-connected layers and uses the migration learning to share the weight parameters of each layer of the VGG16 pre-training model for training. Compared with the traditional image processing methods,AlexNet model,GooleNet model and standard VGG-16 model,the improved VGG-16 model has the highest recognition accuracy of bus congestion status,and the recognition accuracy can reach 96.1%. The loss value of the model converges faster than that of the standard VGG-16 model,and the model is more stable. The experiment proves that the improved VGG-16 model can better extract the characteristics of the crowded state in the bus and solve the problem of the congestion status in the bus.