Key points are not available for this paper at this time.
Road-users classification plays a critical role in efficient transportation planning and management. It is also essential for a number of applications, including electronic tolling systems, roadway design, and intelligent transportation systems. This paper proposed a method to categorize road-users into distinct types using point-cloud data from light detection and ranging (LiDAR) sensors mounted on the roadside (e.g., signal or road poles). Not only motor and non-motor vehicles but also pedestrians were considered as classified objects (i.e., road-users). Firstly, data from the multi-layer LiDAR sensor were pre-processed. Then, six features representing vehicles’ profiles were manually extracted from 3D cloud points as input variables for the classification procedure. Support Vector Machines (SVM), Random Forest (RF), Back Propagation Neural Network (BPNN), and Probabilistic Neural Network (PNN) supervised machine learning algorithms were employed to construct classification models. Based on the pre-defined 8-class scheme within the Federal Highway Administration (FHWA) and two types (i.e., bicycles and pedestrians) defined by authors, road-users were distinguished into ten groups. Compared with the accuracy used by most existing studies on vehicle classification, F1 value, which is popular among the machine learning field, was taken as an indicator to evaluate the performance of classifiers. Finally, the number of features was reduced as an attempt to reduce the dimensionality of sample data for obtaining more accurate classification. The results show that the SVM model utilizing the Gaussian kernel function has the best ability to class road-users with a high F1 value of 84%, and reducing the dimensionality of features can improve the object classification (F1 value increased from 84% to 87.15%).
Song et al. (Wed,) studied this question.