The Research On Detection Algorithm For Front-view Vehicle Based On Monocular Vision

Monocular vision based detection system for front-view vehicle is a system whichaiming to alert drivers by detect front-view vehicle correctly and rapidly. With thedevelopment of machine vision technology, the detection system is widely used in the fieldsof automated driving and intelligent transportation. Major problems of this system are featureextraction and hypothesis verification. Based on the analysis and compare of the researchsituation at home and abroad, this paper put forward new algorithm about feature extractionand hypothesis verification. The research contents are as follows:(1) A shadow feature extraction algorithm based on shadow line scanning and framereduce image filtering is proposed by this paper against missing vehicle easily during thefeature extraction phase. Using the shadow area under vehicle as the priori feature whichextracted in this stage, the algorithm in this pager extract shadow features and zone above itwhich may contains vehicle by fuse a few of shadow lines which locate in same shadow area.Because of the character of the shadow line itself, this algorithm has some advantages such asadaptable to environment, high real-time performance and fairly low missing rate. But it getshigh false alarm rate and waste processing time. In order to ensure the real-time performanceof system, the result set is filtered by frame-reduce image so that most of false alarm could beeliminated. The experiment verified that this algorithm can reduce the misdetection ratesharply.(2) To increase the algorithm’s the accuracy and real-time performance, this paperproposes a verification algorithmic framework which based on voting system. After extractingthe preliminary vehicle information in scene by feature extraction algorithm, system votes oncandidate area detection through a set of simple algorithms just like position estimation,shadow area detection, symmetry test and history information. Then the algorithm classifiesthose candidates by the voting result. The voting system reduces the size of data set which theverification algorithm needs to deal and improves the system’s real-time ability. Because afew of method are collaborating to voting system, the system is more precise for thosevehicles which locate on the side of the image.(3) To solving the problem that the detection system easily be interfered by environment,this page proposes an adaptive shadow threshold method. This method improves system’senvironmental suitability by improved auto-threshold method. In addition, to prevent missing due to environment change, this paper proposes a tracing method based on historicalresult. When the system misses vehicles, this method can fix them by historical information.By these two methods, the system’s environmental suitability can be improved.This paper simulations the measurement system for front-view vehicle in the Windowssystem with Visual Studio2008. The experimental results show that the proposed system ismore stable, accurate and real-time.