Object Recognition And Learning Control Based On Deep Neural Networks For Intelligent Vehicles

As the core unit of the intelligent transportation system,the intelligent vehicle integrates various technologies such as automatic control,artificial intelligence and computer vision,and is a cross-integration product of computer science,pattern recognition and intelligent control and so on.At present,intelligent vehicles have become an important indicator of a country's scientific and technological strength and industrialization level,and have become the focus in both civil and military fields.Environmental perception and motion control are the key technologies to realize intelligent driving.However,the complex and varied driving environment of intelligent vehicles,the high nonlinearity and time delay of the vehicle's own dynamic model make high-precision real-time environment perception and high-performance motion control in complex environments become a challenging problem.In recent years,deep learning(DL)and deep reinforcement learning(DRL)methods have provided a new technical approach to solve this problem.However,the existing DL and DRL methods usually perform parameter optimization based on gradient descent.There are often problems such as the difficulty in generalization and avoiding local minimum values,together with huge training costs due to a large number of search and optimization calculations.As a result,DL and DRL methods for intelligent vehicle perception and motion control have problems of insufficient adaptability and low efficiency.This paper focuses on the environment perception and motion control of intelligent vehicles in complex environments,aiming at the research of fast and high-precision object recognition methods based on deep neural networks,and efficient online learning control method for continuous motion in large-scale state space.The ultimate goal is to reduce the training time cost while maintaining or improving the performance.The main work and innovations include:(1)A novel object recognition method based on deep convolutional features and extreme learning machine(ELM)is proposed,which is named CNN-ELM.The proposed method makes full use of the good feature learning ability of deep convolutional neural network and the superior classification generalization ability of ELM for classification,the advantages of which are complementary.The resultant object recognition accuracy can be further improved without too much increasing the complexity of the model scale.The experimental results on the public database of traffic sign recognition show that the proposed method can achieve 99.40% recognition accuracy.Compared with the best performance method,i.e.multi-column deep neural network(MCDNN),the training procedure of the proposed method is 6 times faster.(2)The influence of color space transformation on the feature learning process of deep convolutional networks is further explored.The perceptual color space transformation is introduced in the deep convolutional network learning process and combined with kernel extreme learning machine(KELM).An object recognition method based on deep perceptual convolutional feature(DPCF)and KELM is proposed,which is called DP-KELM.DP-KELM further improves the discriminability of the features extracted by deep convolutional neural network learning,while reducing the time cost of network training.The comparison results on the traffic sign recognition public database show that DP-KELM can improve the recognition accuracy to 99.54%,and the overall training speed is improved by three times compared with the related CNN-ELM method.(3)Aiming at solving the redundancy and suboptimal problems which may be caused by the random generation of hidden layer network parameters during the learning of the neural network with random weights,the differential evolution(DE)is introduced to optimize the search of optimal parameters.A sparse autoencoder network based on DE is proposed and further integrated as the unit of the hierarchical random neural network(RNN),which is named ESAN.The verification experiments on a variety of standard databases for object recognition show that the recognition performance of ESAN is better than the traditional methods based on the hierarchical neural network with random weights.The training is 2 to 3 or up to 10 times faster than the related deep learning based method such as stacked autoencoder(SAE),deep belief network(DBN)and the like.(4)For the purpose of tackling the online learning control problem in continuous action space with high-dimensional state input,the deep coding network is combined with the cerebellar model articulation controller(CMAC)network to realize the state feature mapping of high-dimensional input,and then the fast adaptive heuristic critic(AHC)algorithm,where recursive least squares TD(?)is used for the value function estimation,is adopted for realizing the online learning of the control policy in continuous action space with images as inputs.The simulation experiment on typical learning control problems show that the proposed method has good data efficiency and learning performance,and can quickly learn satisfactory control policies.(5)A novel lateral control method for intelligent vehicles based on extreme learning machine and deep convolutional features(ICNN-ELM)is proposed.With the expert demonstration information,ICNN-ELM models the lateral control problem of intelligent vehicles in complex environment as the imitation learning problem of continuous action space under high-dimensional state input.Aiming at solving the problem of limited regression fitting precision caused by the insufficient generalization ability of deep neural networks,behavioral cloning technology is used to train deep convolutional feature extractor on expert demonstration data to map the perceptual image to the corresponding control action,where extreme learning machine is used to replace the fully connected layers of the deep convolutional neural network,and the more accurate fitting regression of the control output is realized based on the deep convolutional features.The verification experiments on the Baidu Apollo end-to-end database for lateral control of intelligent vehicles show that ICNN-ELM can achieve a higher accuracy of policy learning while using a simpler network model than the related methods based on deep convolutional neural networks.