A novel multi-hypothesis tracking framework for lane recognition

In the lane marking perception domain, the trend goes to the direction to observe the courses of all lane markings present in an image. Conventional lane tracking systems usually track lane markings under the assumption, that they are parallel. While this model constraint can help to increase the robustness of the system in many situations, it will lead to tracking errors in situations, where the assumption does not hold. On the other hand, if each lane marking is tracked independently, as in normal multi-target tracking systems, the system gets more sensitive to noise, false detections and association errors. We propose a multi-lane tracking system, which maintains the robustness of a parallelism constraint, and also allows to track lane markings following different courses. A novel filter is introduced in this system, it models different lane courses and multiple lane marking offsets in one filter state. Then a multi hypothesis approach is used to assign lane markings to courses and helps to keep the filter robust by deferring the association decision. First results show that a joint estimation of the course assignment and filter variables give a good tracking performance even in challenging scenarios. At the same time the real time running ability is also evaluated.