Research On Airborne LiDAR Data Filtering And Urban Car Object Detection Method

Airborne LiDAR is a new active remote sensing technology. It can directly acquire precision 3D terrain information over large areas. It has been widely applied to many fields, such as obtain high-resolution DEM, urban object extraction and 3D modeling, forest vegetation survey, disaster monitoring, etc.In many applications of airborne LiDAR, it is one of the key steps that classify ground point and non-ground points in the LiDAR point cloud, i.e., the filtering process. Due to the complexity of terrain change and the diversity of surface object features, filtering operations is low degree of automation in actual production. The development of efficient, accurate and automatic filtering algorithm of LiDAR point cloud has been a research focus and facing challenges in airborne LiDAR data processing. Car object detection is also an important research content of intelligent traffic, road construction and tourism planning. The emergence of airborne LiDAR technology provides a new way for automatic detection of car objects.In the above context, this dissertation explored filtering techniques deeply and car object detection methods based on the airborne LiDAR point cloud data. The main research results and innovation points can be summarized as followings.1. On the basis of morphological filtering methods, this dissertation presents an ORF filtering method based on morphological opening by reconstruction without iteration. This method first erodes grid DSM to obtain mark image, and applied geodesic dilation iterately to mark image for achieving the reconstruction process. Then white top-hat transformation is used for nDSM to classify object points with ground points. Finally opening operation is utilized to complete filtering process for quality control. The 15 different complexity test data provided by ISPRS is used in the experiment. Its mean errors of Type I, II, and total error are 3.49%,5.64% and 4.06% respectively. The results show that filtering performance of this method is better than traditional filters.2. To improve largest window size selection in the filtering process of single morphological operators and the degree of automation, a method of urban fusion sequence morphological operators filtering (SMOF) is proposed on the basis of complementary features of morphological operators and urban LiDAR data characteristics of different objects. This method first removed low-outliers and trees, cars, power lines and other small surface features by using morphological opening operation, and then used the morphological gradient to find the edges of large buildings, and finally removed large buildings by using connectivity analysis and binary morphological reconstruction to complete the classificatio of ground and object points. Use 9 urban test data provided by ISPRS in the experiment, its mean errors of Type I, II and total error are 6.90%,3.30% and 5.44% respectively. The results show that its performance is better than that of filtering algorithms of Axelsson and Sithole. It proves the efficiency and accuracy of this filtering method.3. Aimed at the problem of maximum window selection for morphological filtering operation, an automatic filtering method based on iterative multi-scale opening by reconstruction (IMORF) is proposed. In this method, it automatically obtains the erosion window size of reconstruction process by multi-scale method, and solves the misjudgment problem of low object by refreshing the mask image iteratively. The experiments were conducted with 9 urban different complexity test data provided by ISPRS. On the case of uniform basic parameters, its mean errors of Type I, II, and the total error are 3.10%,6.05% and 4.11% respectively. The performance of overall classification and automatic recognition is superior to traditional filtering algorithms.4. CarH method, a car object detection based on airborne LiDAR elevation data is proposed. Based on morphological opening by reconstruction filtering, it completes car detection by the shape feature and elevation attribute of car object. Its mean accuracy and completeness of the three typical experimental regions are 85.2% and 81.3%. It can almost meet the practical requirements.5. On the basis of LiDAR intensity analysis, CarH_INT method, a car detection based on airborne LiDAR elevation and intensity data is proposed. Based on opening by reconstruction filtering, it combined intensity data classification result by Otsu method to obtain high-intensity object points. Then it completes detection in accordance with the shape of car characteristics and its elevation attributes. Its mean accuracy and completeness of the three typical experimental regions were 93.3%and 74.9%. It can almost meet the practical requirements.6. CarH IMG method, a car detection based on airborne LiDAR elevation data and aerial images is proposed. On the basis of opening by reconstruction filtering, it removes vegetation object by NDVI to obtain non-vegetation object, and then completes car detection work in accordance with car shape features and elevation attributes. Its mean accuracy and completeness of three typical experimental regions are 95.9% and 81.9%. It can meet the practical requirements for car object detection under different conditions.