PROBABILISTIC DETECTION AND GROUPING OF HIGHWAY LANE MARKS James H. Elder York University Eduardo Corral York University Lane marks detection Sparsity constraint (inhibition) Detection Refinement (2D Gaussians) Lane Marks Hand-labelling Hypothesis Templates Generation Set of images provided by MTO Graph Generation & Weights Computation Bi-partite graph generation Optimal Matching Locally-connected graph Output Set of chains Introduction Most automated traffic surveillance systems use inductive loops to estimate traffic conditions such as traffic density [3]. The main drawbacks of this technology are the installation/maintenance costs and relatively limited performance. The use of computer vision techniques in traffic surveillance applications enables complex tasks such as vehicle tracking, classification, traffic and motion analysis to be performed automatically at a low cost. Moreover, highway images contain valuable information such as lines, curves and motion patterns that can be exploited. This work is focused on the detection and grouping of lane marks from highway surveillance images. The objective is to use that information in tasks such as camera calibration and image rectification of highways with arbitrary curvature. Taking the logarithm, we obtain our objective function: We seek to maximize this function in order to detect areas of the input image that look like lane marks. Templates. 37 highway images were hand-labelled to generate a set of average templates at 16 discrete angles and 12 vertical regions (to account for the perspective projection). The detector produces both, true and false detections. We use a spatial Greedy inhibition mechanism that eliminates weak detections within a rectangular neighbourhood of the strongest detections. The detections are refined with an unconstrained nonlinear optimization stage. Lane Mark Detection Let Ho and H1 be the hypotheses that an area of the image looks like (or does not look like) a lane mark on the road, and let D be an image patch being analyzed. Assuming an error with a normal distribution and that the elements of D are statistically independent, we compute the likelihood ratio as follows: 1.James H. Elder, Richard M. Goldberg. “Ecological statistics of Gestalt laws for the perceptual organization of contours”, Journal of Vision (2002) 2, James H. Elder, Amnon Krupnik, Leigh A. Johnston. “Contour Grouping with Prior Models”. IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 25, No. 25, June Kastrinaki, V., Zervakis, M., Kalaitzakis, K “A survey of video processing techniques for traffic applications”. Image and Vision Computing 21(2003) Schoepin T.N., Dailey D.J. “Dynamic Camera Calibration of Roadside Traffic Management Cameras for Vehicle Speed Estimation". IEEE Trans. Intelligent Transportation Systems, vol. 4, no. 2, Schoepin T.N., Dailey D.J. “Algorithm for Calibrating Roadside Traffic Cameras and Estimating Mean Vehicle Speed," in IEEE Trans. Intelligent Transportation Systems 2007, pp Patrick Denis, James H. Elder, Francisco J. “Estrada. Efficient Edge-Based Methods for Estimating Manhattan Frames in Urban Imagery”. ECCV 2008, Part II, LNCS 5303, pp , Ministry of Transportation of Ontario: Lane Marks Grouping Weights In order to seek for an optimal matching, we assign weights to the edges E’’ by making use of the proximity and good continuation grouping cues [1, 2]. where is a cues vector relating the i-th and j-th edges and represent distance, parallelism and co- linearity [1]. E is comprised of two disjoint subsets:, where E’ represents the detected segments and E’’ represents potential connections between vertices. We construct the graph G by creating what we call a locally- connected graph. That is, by generating the set E’’ as: where r is a radius obtained from lane marks statistics. Formulation as a Graph Problem We view the lane marks grouping as a graph problem. Let be an undirected graph in which V represents the set of vertices and E represents the corresponding edges. We map the end-points of the k-th detected lane mark (segment) to the vertices. We also map the lane marks to the set of edges E’: We maximize the posterior Where the likelihood term is expressed as: Comments and Conclusions Preliminary comparisons against Hough-based methods show that the proposed method produces superior results. A key advantage is the use of local detection of lane marks coupled with the use of probabilities. The proposed method is expected to produce better results than other published methods especially with curved highways. Our future work plans involve fitting the grouped chains to a model (e.g. Conic), determine vanishing points, perform camera calibration, image rectification and potentially, structure recovery. We constrain the graph G to be bipartite. This enables us to assign the vertices to the sets and such that. Our goal then is to find the optimal matching. The figure to the right depicts an example where a set of chains is determined by the grouping algorithm. These chains could be used for model fitting and for determining vanishing points to perform camera calibration. Set of hypothesis templates Example of refinement via Gaussian model fitting Input Hand labelling The figure to the left shows an example set of detected lane marks outputted by the detector. In general the true positives have stronger responses than the false positives. Grouping Cues: The distributions are learned from a large set of detected lane marks. Then model fitting is performed.. June/17/2010 GEOIDE is funded by the Government of Canada through the Networks of Centres of Excellence program