Intelligent Parking Trajectory Control Method Based On End-to-end Deep Learning

In recent years,with the rapid development of the automotive industry,the concept of the "new four" has gradually become popular in the entire automotive industry.In this change,intelligence,as a core component,has naturally become a hot topic of current research.As one of the key research areas of vehicle intelligence,intelligent parking technology aims to effectively reduce the parking difficulties caused by narrow parking spaces or complex parking environments by assisting or replacing the driver’s operation in the parking process.With the increasing number of car owners,the parking problem is becoming more and more serious,and intelligent parking algorithms have gradually become a key research topic for enterprises and universities.This paper proposes an endto-end deep learning-based intelligent parking entry trajectory method for the challenges of poor real-time and low accuracy of existing intelligent parking models in complex parking scenarios.Its main elements are as follows:To improve the coverage of the data set to the scene and meet the model training requirements,this paper builds a simulation platform based on Pre Scan/Simulink/Car Sim,collects a large amount of environmental data(surround view image data,ultrasonic radar data)and parking control data(steering wheel corner data,vehicle speed data)of the simulated parking scene,and constructs the data set.At the same time,the experimental vehicle,CANoe,combined navigation equipment,and 360-degree surround-view system are used to build a test platform for real parking scenarios,to collect environmental data and control data of real scenarios,and then to build a real parking scenario dataset containing multiple obstacle types,blurred and defaced parking lines,weather factors,and other complex scenarios.Finally,pre-processing operations such as cropping and flipping are performed on the simulated scene dataset and the real scene dataset constructed in this paper to improve the generalization ability and training efficiency of the model.To solve the problems of low parking efficiency and poor parking accuracy in complex scenes that widely exist in current intelligent parking technology,this paper adopts the structure of the parallel combination of convolutional neural network(CNN)and long and short-term memory network(LSTM)to fuse multi-source sensory information and proposes an intelligent parking model based on end-to-end deep learning.The model outputs parking control decisions in an end-to-end manner in real time by extracting spatiotemporal features of 360-degree surround view images and ultrasonic sensor range information in real time.The model consists of four components: an image pre-processing network,which improves the recognition accuracy and training efficiency of the model for complex parking environments by cropping the input images;a spatial feature extraction sub-network,which extracts spatial features and semantic information of the scene using a convolutional neural network(CNN);a temporal feature extraction sub-network,which captures the continuous features of the scene information and control decisions using a long and short time memory network(LSTM);multi-task prediction sub-network,which fuses scene features and coding sequence features by feature cascade method to achieve end-to-end intelligent parking horizontal and vertical control.To further improve the training efficiency and prediction accuracy of the intelligent parking model,this paper fully considers the features of low-speed driving and high localization accuracy requirements in the parking environment and optimizes the sliding window parameters of the temporal feature extraction sub-network to accelerate the convergence speed of the model training and improve the training efficiency.In the model training stage,this paper firstly pre-trains the end-to-end intelligent parking with simulated data sets and then introduces a deep migration learning method to train the model twice with real parking data sets,to improve the prediction accuracy of the model for real parking scenarios and its generalization ability to different parking scenarios.Finally,the model is compared with existing typical intelligent parking models under various real parking scenarios.The experimental results show that compared with the existing intelligent parking methods,the model in this paper has significantly improved parking efficiency and parking performance in complex scenarios.In summary,this paper uses a deep learning method to build a neural network model combining convolutional neural network(CNN)and long and short-term memory neural network(LSTM)in parallel to capture the front-to-back dependencies in scene features and coding sequence data,improve the parking control accuracy of vehicles in complex environments,and better realize the "multiple input-multiple outputs" of the model in parking scenes.In response to the challenge that it is difficult to sense the complex environment of intelligent parking,we propose a method to fuse the information of the vehicle surround-view system and ultrasonic sensors to improve the recognition accuracy of the model;in response to the problems of poor targeting and low efficiency of model training,we propose a method to crop the input image to retain the information of the parking line;in response to the challenges of low efficiency and slow convergence of endto-end deep learning model training in parking scenes,we optimize the model to improve the accuracy of parking control.Finally,a deep migration learning method is used to cope with the limitations of data collection,adapt to the environmental differences between real and simulated complex scenes,and improve the generalization ability of real parking scenes.