Font Size: a A A

Method Of Traffic Road Information Detection Based On Deep Learning

Posted on:2020-09-26Degree:MasterType:Thesis
Country:ChinaCandidate:H B YangFull Text:PDF
GTID:2492306242456804Subject:Computer technology
Abstract/Summary:
With the increasing demand for transportation,and intelligent transportation has become more and more popular,many manufacturers are developing intelligent assisted driving systems.As a special group of people,the disabled need extra care and attention from the society.In this paper,a method of traffic road information detection based on deep learning are implemented,using for a motorized wheelchair assisted driving system for the disabled.Because of the differences between the sense of motorized wheelchair driving and others,current traffic information datasets and methods of object detection have underperformance with low precision and poor stability,which cannot meet the demand of motorized wheelchair assisted driving system for the disabled.To solve this problem,a traffic road information dataset is built and a deep learning method of traffic road information detection based on Faster-RCNN + FPN + ResNet is designed and implemented.A traffic road information dataset with Chinese characteristics from a bike view is built in this thesis.After analyzed the characteristics of extant traffic datasets,raw data is collected from a bike view in Beijing,and annotated with six kinds of object,and finally acquired 14351 images with 40064 annotations.Some data augmentations are implemented to enhance the dataset.A traffic road information detection method based on Faster-RCNN is built and tested with the dataset in this thesis.There are two problems with this method,low detection precision in small objects and mean value.To solve those problems,feature pyramid networks and deep residual network are added and finally an improved traffic road information detection method based on Faster-RCNN+FPN+ResNet is designed and built in this thesis.This method solves the problems of low detection precision in small objects and mean value,and gets mAP=0.8176 on our dataset.Some control groups are tested and analyzed,proving that this method has high precision,strong robustness and reliability,which can meet the demand of motorized wheelchair assisted driving system for the disabled,and has high practical value.
Keywords/Search Tags:Deep Learning, Convolutional Neural Network, Object Detection, Traffic Dataset
Related items