Multi-Class Vehicle and Pedestrian Classification Using Convolutional Neural Network for Traffic Flow and Congestion Composition Analysis
Authors: Robert Kerwin C. Billones, Alexis M. Fillone, and Elmer P. Dadios
Abstract
This paper presents the development of a multi-class vehicle and pedestrian detection and classification using convolutional neural network (CNN) for the analysis of traffic flow and congestion. The study focused on analyzing the traffic flow and volume at different time intervals in a microscopic scale traffic network by decomposing it into eight separate classes of vehicles and pedestrians. Traffic videos in low altitude view T-type intersection (with pedestrian lane and yellow box area), medium altitude view bus stop area, and high altitude view wide intersection are used in the analysis of different traffic flow and congestion scenarios. The CNN model used have a 78.41% training accuracy with 0.6570 loss, and 73.83% validation accuracy with 0.7083 loss for the eight output multi-object classification. The results also showed how each component (class) contributes to the overall road traffic. Private cars constitute about 55-70% of the total traffic volume at any given time, while public utility vehicles (PUVs, jeepneys, buses) only takes approximately 15%. The showed that the implementation of CNN for classification is effective.