Research On Some Key Technologies Of Visual Perception Computing On Intelligent Vehicle

Intelligent vehicles will satisfy the ever-increasing demands of people on automobiles' safety,comfortableness,and convenience,and it can play a very important role in the aspect of improving traffic safety and reducing accident damage,so it has extensive prospect.In key technologies of intelligent vehicle,vision based perception technology is regarded as one of the most potential technologies,which is also a very challenging topic in cognitive computing field.Because of the ambiguity,large data size and interference of image,the ability of computer visual perception is far less intelligent and efficient than human vision system.Moreover,compared with traditional computer vision problem,visual perception computing on intelligent vehicle has to handle with more complex environment,with requirements of hard real-time and high reliability,' so it confronts triple challenges of accuracy,stability,and real-timing.Therefore,research on the key technologies of visual perception computing on intelligent vehicle has important theoretical significance and application value.Aming at the mainstream challenges of visual perception computing,and relying on the Unmanned Ground Vehicle(UGV)evaluation platform of Hunan University,this dissertation mainly research on some key technologies of visual perception computing on intelligent vehicle,by focusing on balancing the property of algorithm in accuracy,real-timing,and robustness to improve the comprehensive performance of algorithms.The main contents and contributions of this dissertation are listed as follows:1.An object detecting algorithm based on improved codebook model is proposed.Considering that the current codebook model cannot conform to their computing features under RGB color space and cannot take both of disturbance rejection and segmentation quality into consideration,this dissertation suggests building codebook model under YUV color space,and establishing Gaussian model for Y component of each code word with the use of Gaussian Mixture Model(GMM)thought,so that the whole code word possess features of GMM.Experimental results show that the proposed object detecting algorithm has distinct advantages in real timing,anti-interference,and detection results.2.A multi-feature object tracking algorithm based on human's visual saliency mechanism.Considering that single feature target tracking has low tracking accuracy and multi-feature tracking can hardly meet the real timing in complex scenes,this dissertation propose to make use of multi-feature particle filter tracking framework to improve accuracy and robustness of tracking,and the discrete particles are filtered by visual attention mechanism.After that,an online update strategy is proposed to reduce repeated computing of features' saliency.Experimental results show that the proposed tracking algorithm is more accurate than single feature particle tracking,and is faster than multi-feature particle tracking method.3.A real-time stereo matching algorithm based on adaptive window disparity refinement is proposed.Aiming at the trade-off between accuracy and efficiency in current local stereo matching,the method adopts different "time-precision" balance strategy in different processing stage,so it has better computing accuracy while guarantees real-time process.Besides,with the use of compute unified device architecture(CUDA),the algorithm is optimized and evaluated in NVidia graphics processor GTS450.Experimental results-show that the proposed method possesses is more accurate and real-timing compared to classic real-time stereo matching algorithms listed on the Middlebury evaluation platform.4.A road recognition method based on vanishing point and principal orientation estimation is proposed.Considering that the currently vanishing point estimation based road recognition algorithm is highly time-consuming,cannot effectively overcome the interference of noise with strong partial textural features,and requires vanishing points to be located inside the image,this dissertation introduces the definition of principle road orientation,so as to reduce computing time and remove disturbance points which have relatively strong partial textural features.Besides,this dissertation proposes a multi-dimension strategy to check whether vanishing points are inside the image,so that the algorithm could handle the situation when vanishing point is outside the image.Experimental results show that compared to the traditional algorithm,the proposed method is more accurate and real-time,and good detection results can still be gained when vanishing points are not inside the image.5.Research on applications of road recognition algorithm on intelligent vehicle system.Relying on an unmanned ground vehicle experimental platform,which is developed by our lab independently,an autonomous driving system model is designed and implemented for evaluating the road recognition algorithm proposed in this dissertation.The system integrates the road recognition method based on vanishing points and principal orientation estimation,and it can detect the road areas in front of vehicles and implement autonomous steering according to real road conditions.Test experiments under real road environment shows that the proposed road recognition algorithm and steering control strategy are effective.