Automatic Detection Of Cars In Mobile LiDAR Point Clouds

With the speedy development of traffic demand, the problem of road safety is becoming serious. The extraction of road information becomes significant in road safety inspection. Precisely feature extraction of cars which take an important part in traffic information is critical. Mobile LiDAR technology provides point clouds with high accuracy. However, the current classification methods possess lower intelligence level. Therefore an automatic car detection method is essential for improving extracting speed and accuracy.The research aims at the automatic detection of cars in mobile LiDAR point cloud which introduces the basic theories including the principle of LiDAR data acquisition, technologies of data scanning and storage and etc. Then analyzes and studies the feature extraction and classification methods. Firstly, the dataset is divided into parts and surface growing algorithm and connected components analysis are utilized to remove the ground surface and group unstructured laser points into components. Then a random selected road part is labeled manually as ground truth and plays a role in evaluation. Features are classified as three groups based on specific characteristics of cars: shape features, contextual features and other features. Contextual feature mainly focus on the position of cars and shape feature considers the size, density, minimum height, height and area of cars the most. Besides, other features as eigenvalues, eigenvalue-based features, RANSAC and reflective information are taken into consideration. Consequently, the feature table would be take advantage of in classification. In the next stage, forward selection and backward elimination are used in selecting features in the feature table. Simultaneously, Bayesian linear discriminant classifier and the linear support vector machine classified cars and static cars respectively. The result is evaluated finally. The foundations are images at the point’s location as well as ground truth. Completeness, correctness and overall accuracy are computed together with the comparison of detecting extent of cars and other objects for quantification analysis.Based on the dataset of Enschede in the Netherlands, the described method of feature extraction and classification are confirmed by a case study. The result shows that features for cars are scientific and feasible; meanwhile feature selection and classification methods are practical and reasonable. Thus effectiveness and operability of the proposed method is proved.