Research Of Vehicle And Pedestrian Detection With Range Estimate Based On Monocular Vision

Efficient and high performance detection on vehicle and pedestrian have been introduced in this thesis,for effectively reducing the injuries,fatalities,and property loss.It also provide range estimate,so that driver can get more information around.By analyzing the advantages and disadvantages of the existing schemes,this paper improves the detection performance of the existing algorithms with low efficiency and low computational efficiency,and proposes a detection scheme based on the 3D probability model.It's widely recognized that targets recognition is important for ADAS,since the user experience is heavily influenced by the result of targets recognition.There are two broadly used methods for improving detection performance: first,find the best feature or integrate several features,so that we can get more accuracy on detection.Second,improve training mode for better performance.In this thesis,the basic Soft Cascade classifier is trained by combining color,grad and orientation histograms with a total of ten feature channels as the feature pool of the initial classifier.In order to speed up the detection progress,image pyramids are changed into training multiple classifiers,so that the amount of computation is greatly reduced.At the same time,the 3D probability model is introduced into the detection process,and the detection performance is further improved by combining semantic analysis,geometric constraints and detector scores.However,targets not move in regular pattern in reality,we cannot predict where they are in future.It's a difficult problem for tracking and particle filter is a good alternative solution deal with this matter.For the particle degeneracy problem of traditional PF algorithm,and only track fixed number targets in the whole process.Author put forward a new modified RJ-MCMC scheme to tracking variable targets,which could update the state immediately when new target add or delete.We can also estimate targets location after stable detection.This thesis mainly studies the assistant driving system under monocular vision,including vehicle and pedestrian detection,multi-target tracking and target range estimation.The experimental results show that the 3D scene probability model combined with classifier detection,semantic analysis and geometric constraint can effectively identify complex targets and ghost targets.Secondly,we design a RJMCMC particle tracking scheme which is suitable for multi-target tracking.It can reduce the drift problem in multi-target tracking.Finally,the range estimation method is improved to reduce the range estimation error,too.