Research Of Vision-based Robust Detection And Recognition Methods For Intelligent Vehicle

Research on the technologies related to intelligent vehicles has important significance to reduce the traffic accident rate,improve the road traffic safety and transport efficiency,ease the driver fatigue level and the situation of energy consumption and environmental pollution.It has been widely concerned by scholars at home and abroad.Specific to the key technology of intelligent vehicle-environment perception,this the-sis researches in-depth the problem of machine vision-based target detection and recogni-tion,the specific research contents and innovations are as follows:(1)A method for stratified camera self-calibration under circular motion with const an-gle inter frame is proposed.The constraints of the infinite plane coordinates between three views are derived from the constraint of circular motion with const angle inter frame.Com-pared to the classic module constraint conditions,the number of possible solutions can be reduced effectively.In addition,a Stratified Iterative Particle Swarm Optimization(SIPSO)algorithm is proposed for the optimization of the infinite plane coordinates.Compared with the traditional method,it can improve the accuracy of the solution by using the sequence of image information in the affine reconstruction step.Experiments prove that the algorithm is robust to noise.(2)A method of camera on-line calibration under urban traffic scenes is proposed.The algorithm is based on the detection of the Manhattan line in 3D scenes,and to estimate the vanishing point of Manhattan direction,and then estimate the parameters of the camera.In addition,the corresponding straight line fitting method is proposed to improve the accuracy of vanishing points estimation for the quasi-Manhattan scene.Moreover,single-frame and multi-frame test schemes are designed respectively.The experimental results show that,although the single frame image can meet the requirements of the camera parameter estima-tion,the accuracy and stability of the parameters estimation can be improved by using the consecutive multi-frames to estimate the parameters of the camera.(3)A traffic sign detection method based on local energy clustering segmentation and region of interests(ROIs)semantic verification is proposed.To segment the image quick-ly,a multiple differential Gauss filter is proposed instead of the Gabor filter to extract the local energy of the image.In order to further verify the ROIs,a method of metric learning with multi kernel embedding is proposed,and the semantic of the ROIs are recognized by combining the color,shape and geometrical features.Experimental results show that the algorithm can effectively eliminate the false positive ROIs,and the algorithm is robust to complex background,illumination and visual angle changes.(4)A traffic sign recognition method based on multi-modal multi-task learning is proposed.In the least squares regression model,two kinds of different structured sparse-induced norms are introduced,one of the norms can be used not only to select modality-of-features but also conduct within-modality feature selection.Moreover,the hierarchical correlations among the classification tasks are well represented by a tree structure,and there-fore,the tree-structure sparsity-induced norm is used for learning the regression coefficients jointly to boost the performance of multi-class traffic sign recognition.Alternating Direc-tion Method of Multipliers(ADMM)is used to efficiently solve the proposed model with guaranteed convergence.Extensive experiments on public benchmark demonstrate that the proposed algorithm leads to a quite interpretable model,and it has better or competitive performance with several state-of-the-arts methods but with less computational and memo-ry cost.(5)A robust pedestrian detection method by combining high resolution and low resolu-tion features is proposed.The algorithm uses linear transformation matrix to map the feature of high and low resolution sample to the same feature space,and to learn the best detection classifier in this space,and the shared detector in this space is learned jointly based on the linear SVM model.Meanwhile,the maximum mean discrepancy(MMD)is considered to minimize the distribution mismatch of the transformed high-and low-resolution features.Then the robust detector and the transformations are learned jointly by minimizing both the structural risk functional and the distribution mismatch between the mapped high-and low-resolution samples.In addition,we show that our proposed learning framework can be converted to a standard Multiple Kernel Learning(MKL)problem,which is convex and hence the global solution can be guaranteed.Experiments for the pedestrians taller than 30 pixels on the Caltech Pedestrian Benchmark show that our method is effective for the low resolution pedestrian detection tasks.