Study On Vehicle Forward Collision Warning System Based On Monocular Vision

The problem about traffic accidents is one of the worldwide public safety issues which is getting worse continuously.According to the statistics,main cause of traffic accidents is the driver rather than motor vehicle.90%of these accidents can be avoided if there’s a 1.5 s warning just before the accident.Thus the development of techniques which could perceive the surroundings during driving and make corresponding measures to protect the driver when the motor vehicle in danger is a topic continuously gains attention.Based on computer vision research achievements,a new Forward Collision Warning System which could effectively protect the driver during driving was developed for real-time urban traffic environment.With the infomation provided by the sensor camera,drivers could perceive and estimate the potential dangers while driving.FCWS would warn the driver of the potential dangers before they actually happen.If the driver who receives the warning does not do anything about it,the FCWS would take over the vehicle and take actions to avoid accidents.The research content in this thesis is listed below:1)Foreground segmentation method for vehicle forward collision warning system is researched.This thesis presents an improved foreground segmentation algorithm based on objectness feature.The prospect of the traditional segmentation algorithm is much affected by vehicle speed detection results,it is difficult to meet the dynamic environment of vehicle detection system with the requirements of the target accurately.In addition,the traditional algorithm is very sensitive to the changes of illumination thus need post-processing.Therefore in this thesis,using the objectness property characteristics is proposed to calculate the weight of the general characteristics of the objects in the image area,through sorting each region of the score to predict whether the region may contain objects,so as to avoid the change in the dynamic environment of foreground segmentation effect.In this thesis,the candidate proposals obtained are divided into three categories,and the objectness of partial object proposals are used to enhance the weak weight of whole object proposals,reducing the cost of calculation.The Fast R-CNN deep learning network is extended to meet the requirements of detection proposals training and can be combined with subsequent training applications.A small amount of time is used to obtain a foreground window that containing the vehicle,providing plenty of time for the subsequent processing of the front collision warning system.2)Vehicle detection method for vehicle forward collision warning system is researched.Based on features fusion model,this thesis proposes an algorithm for vehicle detection.The traditional vehicle detection technology does well in recognizing the vehicles under the static environment.However,in a dynamic environment the emergence of new vehicles will give rise to false detection while using traditional technology.This thesis presents forward a model based on feature fusion that uses Haar features and Adaboost cascade classifier to predict undetermined area image carries on the preliminary classification.We use HOG feature and SVM classifier to verify whether the pending area contains vehicles.Experiments show that the algorithm not only avoids the false detection caused by the rapid change of background,but also ensures the real-time performance.3)In this thesis,an emerging-occlusion vehicle tracking algorithm based on corner is proposed.The conventional method of corresponding to a certain extent can solve the problem of multiple vehicle tracking,but it can’t distinguish between block each other and split the multiple vehicles.The proposed tracking algorithm based on corner point is used in vehicle sheltered area and training corner point information is used to the classification of the area of the current block corner,so as to resist each other shade and no effective tracking of multiple vehicles before the split,its main including two parts content:corner feature extraction and classification.4)A general introduction of Forward Collision Warning Systems is made in the first place.This part includes the progress made in the area so far and an illustration of the general requirements for the designation.Camera calibration and distance measuring method for vehicle forward collision warning system is researched.This thesis proposes to use a combined camera with 3D laser radar calibration approach to add 3D distance information into visual images.At the same the speed provided by the vehicle system and the distance between self-vehicle with forward vehicle can be calculated to predict the time to collision(TTC).A demonstration of our system is also presented.Finally,this thesis performs experiments as the experimental platform in real urban environment.The results show the correctness and effectiveness of designed FCWS that can ensure the safety for the driver while driving.