A Generic Temporal Integration Approach for Enhancing Feature-based Road-detection Systems

Visual feature-based approaches for detecting the drivable area on unmarked streets and roads were introduced in recent years. Although the accumulated results are promising, the detected street segments often contain holes and show a detection performance that strongly varies in time depending on environmental conditions. This paper presents a real-time capable approach for improving the road detection results for this type of state-of-the-art systems by adding a generic postprocessing step. Our proposed architecture removes the drawbacks of said systems using a temporal integration approach based on the bird´s eye view. In order to test the proposed approach, one typical visual feature-based road detection system was implemented. Still, the used road detection system can be exchanged with any other state-of-the-art system. Evaluation results computed on inner-city data show that this approach is an important enhancement for all visual feature-based road-detection systems. One of the used streams and corresponding ground truth data is accessible on the internet for benchmark testing. The proposed approach is a crucial step toward robust road detection in complex scenarios that allows building high-level applications, as, e.g., active collision avoidance or trajectory planning, based on vision as the major cue.