Research On Vehicle Intelligent Ranging Algorithm Based On Target Detection And Monocular Vision

The intelligent ranging algorithm based on object detection and monocular vision studied in this paper is an important part of modern autopilot field,which mainly includes deep learning technology,monocular vision technology and so on.With the continuous improvement of computer hardware level,deep learning technology has been developed rapidly in recent years,the depth of convolution neural network is increasing,and the scale of parameters is increasing.The application of various optimization algorithms makes the training time of convolution neural network shorter,the accuracy rate is improved,and the operation is more efficient.Now the technology has been gradually applied to the field of self-driving,unmanned field.In this paper,an improved convolutional neural network target detection algorithm and a monocular vision ranging algorithm are used to build an intelligent ranging system.In this paper,the adaptive gradient descent algorithm is used to reduce the fluctuation of the network during training,and the time of model fitting data is shortened.By using Dropout random deactivation algorithm and weight regularization,the generalization ability of the model is improved,and the accuracy of the model is improved effectively.train the network through big data,and then combine the obtained network model with monocular vision ranging technology,and install the system in the on-board embedded system.In view of above content this article will carry on the research from the following several aspects:Convolutional neural network construction: according to the special requirements of vehicle driving auxiliary system,such as the high real-time processing of information in the event of emergencies,the recognition of vehicles and pedestrians on the road of the main object accuracy is high.This paper presents a neural network with 16 convolutional layers.It can not only meet the real-time requirements of neural network processing information,but also meet the requirements of network accuracy.And can be installed in most embedded systems.Optimization of network parameters: selection of activation functions,mainly determined by a large number of simulation experiments in playground-tensorflow experimental systems.Relu function can accurately fit the training data,get a good network model,and in the test data compared to the other activation function accuracy is the highest,so this paper chooses the Relu function as the activation function.The processing of network model overfitting: if the network model over-fit the training data,then the generalization ability of the model to the new data will be weakened,and the accuracy will be reduced.Aiming at the above problems,this paper introduces an improved weight regularization algorithm in the convolution layer and a random deactivation algorithm for Dropout neurons in the full connection layer,which makes the accuracy of the network model significantly improved.BP algorithm optimization: the weight update of convolutional neural network is carried out by backpropagation.in the process of backpropagation,the adaptive gradient descent method is proposed,which can effectively improve the fitting speed of the network model,and can make the network model oscillate at the lowest point of the loss function,thus improving the accuracy of the network model.After optimizing the convolution neural network and target detection algorithm,the classification and location accuracy of the algorithm meet the requirements of application in autopilot system,and the ranging error is controlled within a reasonable range.