Real Time Detection and Classification of Arrow Markings in Urban Streets

For highly automated driving in urban regions it is essential to know the precise position of the car. Furthermore it is important to understand the surrounding context in complex situations, e.g. multilane crossings and turn lanes. To understand those situations there is not only the task to detect the lane border, but to detect the painted information inside the lane. The paper is facing and evaluating two methods to classify this additional lane information. Therefore the images from five cameras mounted around the car are used. Four of them with fisheye lenses. The methods have in common, that the input images are transformed into a bird view projection. First introduced method is to extract contours from the transformed images and collect geometrical features and Fourier coefficients. The second introduced way, is to calculate histograms of oriented gradients and use it as input for the classification step. Both classification approaches are implemented and evaluated as multiclass and single class detectors for each arrow type. Furthermore, the classification results from a support vector machine and random forest were faced for this classification problem. The results from the multiclass detectors are evaluated and presented in form of confusion matrices. With the introduced approaches a high detection confidence could be achieved, proofed with validation datasets and in practical use.