The Optimal Parameter Combination For Vehicle Detection From WorldView-2 Imagery

Vehicle detection and traffic flow parameters extraction from high resolution imagery have been popular research topics in recent years. Object-based image analysis, with the aid of commercial eCognition (R) software, has become a widely implemented method for vehicle detection and identification. In the process, the determination of optimal segmentation parameters is of crucial importance for the accuracy of vehicle detection and classification. In this study, we obtained the optimal parameter combination for vehicle detection by a "trial-anderror" strategy using road subsets as sample clipped from WorldView-2 imagery. A better vehicle detect result would be obtained easily in 0.5m resolution images when the scale value was 30, the shape value was set to 0.4, and the compactness value was kept constant as 0.5. At the same time, an obvious tendency was also revealed that the segmentation accuracy of vehicles tended to descend gradually holding the scale value constant with the increasing of shape value, and the accuracy descended sharply when the shape value was larger than 0.5; it was also descending generally holding the shape value constant with the increasing of scale value. The experiment results can be applied directly to the vehicle detection from the same resolution images with similar vehicle size and would also be a useful reference for quick setting of original parameters in other conditions.