Mysliwetz and Dickmanns 1987: An efficient distributed processing scheme has been developed for visual road boundary tracking by ‘VaMoRs’, a test-bed vehicle for autonomous mobility and computer vision. Ongoing work described here is directed towards improving the robustness of the road boundary detection process in the presence of shadows, ill-defined edges and other disturbing real world effects. The system structure and the techniques applied for real-time scene analysis are presented along with experimental results. All sub-functions of road boundary detection for vehicle guidance, such as edge extraction, feature aggregation and camera pointing control, are executed in parallel by an onboard multiprocessor system. On the image processing level local oriented edge extraction is performed in multiple ‘windows’, tightly controlled from a hierarchically higher, model-based level. The interpretation process involving a geometric road model and the observer’s relative position to the road boundaries is capable of coping with ambiguity in measurement data. By using only selected measurements to update the model parameters even high noise levels can be dealt with and misleading edges be rejected.

1. INTRODUCTION
At the UniBw München a test-bed vehicle for autonomous mobility and computer vision has been in operation for about two years by now. Besides system integration the main effort initially has been focused on fundamental aspects of vehicle modeling, control and road parameter estimation for high speed visual guidance [1, 2]. Extended autonomous rides on a well visible highway lane were demonstrated at speeds of up to 96 km/h in August 1987, proving the potential of applying dynamical models and estimation methods to real time image sequence analysis [3].
   The emphasis of ongoing work described in this paper is on improving the road boundary detection and tracking process to achieve reliable results under adverse lighting or contrast conditions. Multiprocessor implementation aspects are also covered in this context. The approach presented employs a hierarchical, model-based processing scheme to cope with high noise levels. Multiple local edge extraction on the low level produces redundant but ambiguous feature descriptions. At a higher level features are selected from a set of potentially ‘good’ candidates and then used to update the model parameters. With the model serving as a basis for the selection step and as a guiding mechanism for feature extraction, only local image operations are necessary.
   The various techniques being investigated for road boundary tracking can be roughly categorized into a) edge based using intensity images b) region based using intensity images and c) region based using color images [4, 5, 6]. The approach described here is of the first type as far as the image processing level is concerned. Though this method might be considered the most susceptible to disturbances from shadows or low contrast edges it will be shown that in combination with a simple model that approximates spatiotemporal continuity constraints of observer motion, it is robust and efficient at the same time.
   Basically the principle of incrementally updating an internal scene- or process- model is applied. Computing reference feature positions based on the current model parameters and then taking the weighted difference between reference and actually measured positions in the next frame is used to propagate the model parameters. Having a reference through the model also allows to establish a feature selection mechanism to derive better measurements from a set of noisy and ambiguous data which in return result in better estimates of the model parameters. The current implementation does not (yet) employ an explicit extrapolation step by a dynamic model of vehicle motion. However, a fast processing cycle rate of 30 Hz keeps the image matching problem highly constrained and is essential here for tracking dynamics and noise robustness. As will be shown, already a rather simple and straightforward model is capable of compensating for ambiguous measurements or even bridging partial gaps in the sequence of measurements.