Vehicle Location And Identification Based On Deep Learning And Reinforcement Learning

The technology of vehicle location and identification is a one of the key technology in intelligent transportation system.It is the essential prerequisite of the target vehicle searching,the vehicle speed monitoring and the vehicle tracking.The research object of this paper is how to employ the idea of deep learning and reinforcement learning into vehicle location and identification in a static image.The target of identification is to determine whether vehicle existing and vehicle belong to which category,such as car,bus and truck.The vehicle can be located by using the position information of pyramid image technology.This paper is organized as follows:First, the two major network models are analyzed:deep belief network and convolutional neural network.The structure and principle of the two models are presented and training algorithms are analyzed in detail: gradient descent algorithm and greedy algorithm.The two models are used for the simulation experiment of vehicle identification.The simulation results indicate that deep network is better than traditional shallow network in vehicle identification and also further proof that convolution neural network performance well on the vehicle database in this article than deep belief network.Second, the idea of reinforcement learning is used to the convolutional neural network.The combination model of convolutional neural network and reinforcement learning is applied to vehicle identification.The theoretical basis of the combination model and Q-learning algorithm are analyzed.In view of the core of Q-learning lies in the action selection strategy of the Agent,we proposed the idea of vehicle categories mapping for the actions.So we can training the deep neural network by using the action-based selecting evaluation Q value instead the gradient of output error to spread in the convolutional neural network.For dealing the issue at the end of the training procedure in convolutional neural network which due to the small sample proportion of misclassification and lead to a small influence to update network weights,we proposed a reinforcement learning strategy which dynamically adjust the training set based on the misclassified samples.The simulation results indicate that the proposed reinforcement learning strategies can effectively improve the performance of the vehicle identification and is conducive to the training of deep neural network.Finally, based on the theory and experiments designed to achieve A practical system for vehicle location and identification is designed and implemented based on the above theory and experiments.The system uses the multi thread processing technology and provides functions such as interactive interface,raining and testing of the deep network, vehicle location and identification,etc.The results indicate that the system can identify cars, buses, trucks better under the road environment.The design requirements can also be satisfied.