Hybrid of gaussian and wallis filter for urban area characterization of LiDAR images

Abstract

Visual clarity of aerial LiDAR image textures with integration of LiDAR points is critical for accurate multi-features extraction in dense urban landscapes to provide better input images. The urban landscapes are composed of heterogeneous building footprints which caused incomplete building cues when extracted. Besides that, these images include textures of rough surface such as building concrete and asphalt (road) which often suffer from scene reflectance. Due to surface reflectance, an input image must be corrected by applying noise filtering method to the image before extracting features. This research proposed a unification of Canny Edge Detection technique to visualize image information of isolated building footprint edges. Based on a dataset of Bukit Kanada, different techniques such as Canny Edge Detection, Individual Tree Detection (ITD), Semivariogram and Self-Organizing Road Map (SORM) were used for extracting features based on four characteristics of an urban area which are buildings, trees, artificial and natural ground. By combining Wallis and Gaussian filter, image enhancement was performed on input aerial LiDAR images that contained noise to improve the interpretability information of images. Besides that, to extract heterogeneous building footprint with less errors, a unification of Canny Edge Detection technique with Hysteresis, Boundary-based Segmentation, Hough Transform and Chamfer Matching was performed. Skewness and kurtosis statistical results showed that the proposed technique produced better surface textures for features extraction of an aerial LiDAR image. In addition, there are substantial improvements of urban multi-features extraction from the enhanced image in terms of better surface clarification and noise reduction performance. The standard performance measures of the newly enhanced image in terms of completeness, correctness and overall quality are better and have more accurate building detections with a smaller number of omission errors as compared to the other existing techniques