Robust Optical Flow For Driver Assistance System

The Research of Driver Assistance System (DAS) becomes more and more popular in computer vision domain in recent years. DAS endeavor to provide a safer driving environment and reduce avoidable accidents:there is intense interest in them, because even a small reduction in accidents has enormous social and economic benefits. DAS use some fixed cameras to obtain the information for the drive environment in virtue of simulating human vision mechanism, processing the information to detect and determine the potential danger. For smart, safe vehicles, DAS must have three characteristics:real-time, robustness and practicability. Real-time refers to the constraints on processing time induced by the vehicle's speed. Robustness means that the intelligent vehicle can adapt to complex surroundings. Practicability is concerned with consumer acceptance of DAS for its effectiveness. To this end, the study of new algorithms to meet real-time sense judgments and robustness performance become to the key problems for nowadays computer vision domain.Optical flow method uses image pixel grey values and their changes in the time domain to calculate the motion of each pixel. It computes a flow field representing the motion of pixels between two consecutive image frames. Optical flow is very important for DAS, since it reveals abundant 3D target information. The classical optical flow (Horn-Schunck Optical Flow) method uses two basic assumptions, one is the intensity consistency assumption (data constraint) and the other is the smoothness constraint (regularization). Unfortunately, in real-world scenes encountered while driving, the intensity consistency assumption is often violated. Furthermore, the smoothness constraint does not allow for displacement discontinuities, and it does not handle outliers in the data term robustly. Therefore, Horn-Schunck model is not practical for DAS.In this paper, we employ a coarse-to-fine warping a novel preprocessing procedure-high boosting-to realize the robustness. We show experimentally that our method is more robust than the Horn-Schunck model and the original TV-L1 while preserving discontinuities; the high boost filter is more robust to illumination variance than other filters such as median ones and Sobel ones and produces similar results to Wedel et al's structure-texture decomposition at far less computational cost. Details include the following four aspects:1. Research the preprocessing of images. Filtering through the low-frequency and high frequency filters to obtain structure and texture images, then compare and analysis the optical flow of them and the original images, we novelty imply the high-boost filter in our DAS, not only improved the robustness of optical flow calculation in illumination variance environment, but also retained the structure information in the low-frequency images.2. Research the image pyramid for preserving discontinuities. Scale-space solutions give well separated and contrast preserving features at different scales. Although pyramid processing reduces an image to many scales, computation time is increased by a factor of only 4/3 and it handles the large disparities efficiently. Whilst grey values can simply be averaged, flow vectors need to be scaled with the scaling factor between the pyramid levels to yield valid displacement vectors on every level.3. To speed the computation up, we replaced the nonlinear intensity profile by the first order Taylor approximation to linearize the problem locally.4. To improve visualization of the flow vectors, we encoded optical flow into HSV color space.Finally, the summary of our work has been given, and we put forward the future work.