Video Target Detection And Tracking For Vechicle Driving Assistance Systems

A vehicle driving assistance system(VDAS)can improve the driving safety and reduce the traffic accident rate.As the key technology of the driver assistance system,environment awareness has got great attention and rapid development in recent years.However,in the driving environment,lane markings usually have complicated structures,vehicles have different types,lighting lillumination changes dramatically and the background interference is complex.What's more,drivers' driving skills and attention degree differ in thousound ways.All of them create great challenges for the fast,accurate,compressive-orientated design of driver assistance system.To solve the problems mentioned above,this disseration should research on how to use the camera to perceive the lane markings,front vehicles,and the driver face,which provides general detection and tracking methods of the typical video objects for the design of VDAS such as lane departure warning system(LDWS),forward collision warning system(FCWS)and driver fatigue monitor system(DFMS)and so on.At the same time,this disseration build some experiment platforms using Windows OS and embedded Linux OS to validate the recognition of lane and face.The main research contents and contributions are as follows:(1)Multi-sacle match filter-based lane detection scheme of constant false alarm rate was provided.From the perspective of producer/consumer model of the whole system,the necessity of matched enhancement is put forward,and multi-scale matched filtering method is proposed according to the different width characters of lane markings,then an adaptive dynamic threshold is set based on local noise estimation form the improved mathematics model of constant false alarm rate.Sbusequenly,lane markings are fitted using random sample consensus,and lane parameters are filtered by the online lane parameter database.Through comparasive analysis on road experiments,the adaptability of this method to various lighting and weather conditions was tested,and the robustness to recognize lane marking of different width was verified under low contrast scenes,such as sunshine and shadow disturbed strongly and lane markings worn badly.(2)Bayesian probablity decision-based vehicle boost learning method(Bayes Boost)was proposed,and front multi-vehicle tracking method based on compressive sensing on multi local PCA/ICA features.As for front vehicle detection,based on the relatively large scale image database of vehicle rears,the classification and detection performance of original cascade Ada Boost algorithm was improved using the Bayesian probablity decision.As for front vehicle tracking,in order to improve its adaptation on occlusion,variation and sacle change,the relatively complete sets of PCA/ICA features are extracted just in four local regions in vehicle rear images,in which PCA features is corresponding to the low frequency features of rear images and ICA features corresponding to the residual images after extracting the PCA features.Compressive sensing on the PCA/ICA features improves the features' sparse characteristic and also makes the features subject to Gaussian distribution,laying a good foundation on deciding the class probability of child windows using the Bayesian posterior probability rule.A step-by-step finely fast search method is custom-built,which makes the front multi-vehicle tracking into reality just using single object tracking algorithm.Through comparative analysis with some similar tracking methods,the method's performance on front vehicle detection and tracking are tested on the complex traffic scenarios.(3)Driver face detection based on deep convolution representation was employed,and a Discrete Cosine Transform(DCT)perceptual hashing-based face tracking method was prenented.Since driver faces have great individual difference,it is necessary to use deep convolution network to unsuperivisedly learn and represent the driver face features.Through the deep train on the large scale face database,a face detector of good performance is produced.In the face tracking stage,the main low frequency information of face region are extracted by DCT method,and these DCT low frequency information are encoded by the perceptual hashing method,which is used to be applied in the image similarity retrieval domain,making the face tracking method be suitable to various lighting conditions.The experiments on the typical face testing videos and our driving videos show that,the proposed method can stably track driver facesin the dynamic lighting and position-change scenes at the video frame rates.(4)A serial of validate and design solution based on multi-paltform like Windows and embedded Linux OS are presented.A lane detection validate platform is built using a ARM-based architecture,in which the real traffic video played on the LCD is captured through the camera of high dynamic range.What's more,the embedded hardware configuration and some soft drive programs of key interface are presented.The experiments shows that the lane detection and tracking method is of good performance on detection,tracking,and has realtime processing rate.At the same time,a fast driver face detection and tracking program based on driving video stream is presented using the multi-thread technology on Windows operation system.