Research On Vision Parking Algorithm Based On Convolutional Neural Network Major:Communication And Information System

Recently the research on autonomous vehicles have witnessed more and more attention and made great progress,which provides more possibilities of the active safety and intelligence of autonomous vehicles.The environmental complexity of parking and the application possibilities can be achieved more simply than other driving technologies.However,the existing parking assistance is mostly based on multiple ultrasonic or multi-radar to perceive the parking area.The high application cost makes it not fall on more car models.With the rapid development of computer vision,the lower cost camera has been developed in the field of parking.As the first step in the three parts of the parking system which contains perception,decision and control,the perception is to detect and identify the parking line in the object area during parking.The traditional image object detection algorithm has low recognition accuracy for complex scenes and depends largely on the setting of parameter thresholds.Deep convolutional neural network has better performance,greater application potential and research value.Although the speed of parking is relatively low,there is a high requirement for the real-time speed of the detection algorithm,so it is especially important for the optimization of the algorithm.The work of this paper is mainly to study the parking visual aid algorithm,propose a CNN based parking space detection algorithm and the algorithm is validated on the PC.The main research contents of this paper are:(1)The object detection algorithm based on convolutional neural network.In this paper,the region-based detection algorithm(two stage framework)and the unified pipeline detection algorithm(one stage pipeline)have been introduced in detail.A small size parking space dataset of a single scene have been made,and the corresponding experimental comparisons are made for both types of algorithms.The accuracy is further improved by data augmentation,structural optimization,and parameter tuning.(2)Research on model compression and acceleration based on deep neural network.Through the research and analysis of the existing algorithms,firstly using the algorithms applied to the classification problem to compare and optimize.The difficulties of the application and the training model skills have been analyzed.Finally,the compact small network method has been selected to further optimize,which achieves to compress the object detection algorithm model to ensure the accuracy and improve the speed of forward inference.(3)Research on multi-angle object detection algorithm.Mainly adopted three ideas for research and simulation.Firstly,the image post-processing research optimization of clustering and morphological operations is performed on the detection results of the object detection algorithm.Secondly,improve the multi-angle object detection algorithm combined with natural word processing,and combine detection algorithm CNN and RNN.Thirdly,Input the object detection algorithm's result into another landmark detection model,set 12 landmarks for the four corners of the parking space to detect the parking space contour.Providing feasibility for a rich inspection scenario,providing image post-processing for multiple situations and end-to-end detection schemes.