Robust vehicle recognition algorithm using visual saliency and deep convolutional neural networks
1.Automotive Engineering Research Institute, Jiangsu University, Zhenjiang, Jiangsu 212013, China; 2.School of Automotive and Traffic Engineering, Jiangsu University, Zhenjiang, Jiangsu 212013, China
Abstract：To solve the problem that the traditional manual feature based machine learning algorithms were unable to satisfy robust vehicle recognition requirement in complex traffic environment, a vehicle candidate extraction method was proposed based on visual saliency theory. In the framework of deep learning, a robust vehicle recognition algorithm was also proposed based on deep convolutional neural networks (DCNN). The DCNN was constructed with two hidden layers of filtering layer and pooling layer. The raw pixels of gray images were set as input and trained by stochastic gradient descent algorithm. The whole networks were output by a fully connected layer. The KITTI library was chosen as test database to complete the vehicle recognition experiments. The results show that the overall detection rate of the proposed method reaches 98.13% with false detection rate of 1.77%, which is better than that of existed vehicle recognition methods.
2015, Vol. 36 
