Research And Realization Of Image - Based Driving Active Safety Front Anti - Collision

Collision avoidance of preceding vehicles is an important research direction in the field of active safety driving,the core problem is the detection,recognition and tracking of forward vehicles.And the main processing operation of detection,recognition and tracking of preceding vehicles is performed for the rear face of vehicles,this paper mainly focuses on the detection,recognition and tracking of the rear face of vehicles,the main contents are as follows:(1)A coarse positioning method of the rear face of vehicles based on the shadow underneath a vehicle is presented.In this paper,a modified Gaussian mixture model clustering algorithm is used to cluster targets in traffic images such as the roads,lane lines,vehicles and shadow underneath a vehicle,shadow threshold is calculated according to the mean and varance of the Gaussian shadow model,thus both the shadow image and the intersecting lines of the vehicle shadow and the road are generated,finally the initial position of the rear face of vehicles is located by merging the information of the aspect ratio of the vehicle and the shadow lines.This method can not only effectively locate the rear face of a vehicle but also adapt to the change of the intensity of light at different times during the day and the interference of the external environment.(2) This paper studies two different rear feature extracion methods of a car of Haar-like and HOG,proposes an improved vehicle verification method based on Haar-like and Adaboost,this method can effectively eliminate false alarms generated in the coarse positioning stage.This article uses Adaboost and SVM classifier analyzes the differences and the vehicle identification result of the two feature extraction methods.In order to better highlight the horizontal and vertical edge features of the rear face of vehicles increase the recognition ability and robustness of the classifier,this paper uses gradient image to extract Haar-like features and incremental training methods of Adaboost classifier.(3) This paper studies two kinds of vehicle tracking methods of the rear face of Kalman filter and particle filter.it is used for tracking the rear face of vehicles obtained in the last stage.Firstly this paper introduces the implementations princile of the two vehicle tracking methods and analyzes their advantages and disadvantages in the realization process.Then the vehicle tracking ability is tested on the images captured by the vehicle mounted camera video and the tracking result is also analyzed.Vehicle tracking result based on particle filter is better than the method based on Kalman filter.However,as the number of particles increases,the computational complexty becomes very high,the resampling stage will cause the loss the of particle diversity and effectiveness.(4)An improved vehicle tracking method based on TLD algorithm is achieved.Vehicle tracking method based on TLD algorithm which combines tracking module and detection module,uses P-N learning module continuously to update the feature points of tracking module and the target model of detection module,is possible to realize stable and reliable tracking of the preceding vehicle.This article uses objectness Estimation method instead of the sliding window method used in the original detection module,the modified method speeds up the target detection speed and better meets the needs of real-time system without the loss of detection accuracy.