Researches On Key Image Processing Technologies For Vehicle Active Safety

As an important part of advanced vehicle technology, the vehicle active safety technology plays an important role in the security of human life and property, the efficiency of transportation improvement and the energy consumption reduction. Due to accurate, intuitive, flexible, and without interference with other visual sensor, the vehicle active safety systems using visual sensor have drawn a lot of attentions from many auto manufacturers and researchers. Based on classical multi-resolution analysis method and recent sparse representation theory, this dissertation investigates some key problems of image super-resolution, lane detection and visual object tracking in vehicle active safety systems. Some novel image processing methods based on multi-resolution analysis method and sparse representation theory are proposed. The main contributions of this dissertation are as follows.1. A novel image super-resolution scheme based on hybrid multi-resolution analysis is proposed. Since there are smoothing edges and ring artifacts in the high-resolution image reconstructed by the interpolation method in the wavelet domain, firstly, this dissertation investigates the feature expression ability of different multi-resolution analysis methods to different morphological components mixed in an image. This dissertation finds that the cartoon component and the texture component can be well represented by the stationary wavelet transform and the nonsubsampled contourlet transform respectively. Then, a multi-morphological image super-resolution method based on hybrid multi-resolution analysis is proposed in this dissertation. Therefore, the cartoon component in high-resolution image is reconstructed exactly, while the edges of texture component in high-resolution image are well preserved. Furthermore, the ring artifacts along edges in high-resolution image are eliminated, since the stationary wavelet transform and nonsubsampled contourlet transform is shift-invariant. The experiments over intelligent transport systems surveillance images and natural images indicate the effectiveness of the proposed method.2. Two novel multi-morphology image super-resolution algorithms via sparse representation are proposed. Firstly, in order to take full advantage of the priors between the low-resolution images and high-resolution images, a multi-morphological training sets construction method and a multi-morphological dictionaries training model based on sparse representation theory are proposed, which can obtain the natural characteristics of the image using a few atoms from the over-completed dictionary. With assumption that the low and high resolution images have the same sparse coefficients in relation to their own dictionaries, a novel multi-morphology image super-resolution algorithm via sparse representation is proposed. Secondly, in order to combine the sparse representation theory and the feature of natural images, this dissertation investigates the different features of different morphological components mixed in natural images. With the singular value decomposition dictionary learning algorithm and the orthogonal matching pursuit algorithm, a multiple dictionaries training framework is presented to effectively learn multiple morphology dictionaries. Then different morphological components are combined with reasonable morphological regularizations, for example, the texture component regularized by the non-local similarity and the cartoon component regularized by the total variation. As a result, another novel sparse representation algorithm with reasonable morphological regularization is presented for single image super-resolution. Extensive experimental results on various intelligent transport systems surveillance images and natural images validate the superiority of the proposed algorithms in terms of qualitative and quantitative performance.4. A lane marking extraction algorithm using orientation and vanishing point constraints in structured road scenes is proposed. In order to deal with various illumination conditions, clutter shadows, and other interferences from the external circumstances at the same time, firstly, a road surface region adaptive segmentation method is presented to eliminate the interferences from the non-road surface circumstances. Furthermore, most of interferences from the external circumstances are eliminated by Laplacian filter. Secondly, the line segments detected by line segment detector are applied to represent the structural information of lanes. Thirdly, non-lane candidate markings are removed by clustering analysis methods with orientation and vanishing point constraints. Finally, the lanes are extracted from the remaining candidate lane marking in very high accuracies. Experiments on complex structured road scenarios in urban streets show that the proposed algorithm can effectively filter the interferences from complex structured road scenarios, and then accurately capture the left and right straight lane marks close to vehicle with low average false detection rate and average missing detection rate. 5. A sparse representation algorithm with super-pixels for visual object tracking is proposed. Since trivial templates in tracking method using sparse representation can not accurately address the occlusion problem, this dissertation incorporates super-pixels into sparse representation to effectively and flexibly represent the structure information of image. Then this dissertation utilizes super-pixels masks of the reconstructed object image to accurately detect occlusion rates, and then the proposed algorithm does not require a lot of trivial templates which will lead to high computation cost and inaccuracy detection of occlusion. Therefore, a sparse representation algorithm with super-pixels for visual object tracking is proposed. Experimental results on challenging image sequences show that our algorithm is able to deal with heavy occlusion, and effectively track the target through scale and appearance changes.