A Real-time Vehicle Detection for Traffic Surveillance System Using a Neural Decision Tree

Traffic surveillance system (TSS) is an essential tool to extract necessary information (count, type, speed, etc.) from cameras for traffic monitoring in many metro cities. In TSS, vehicle detection plays a pivotal role as it is a vital process for further analysis such as vehicle classification and vehicle tracking. So far there has been a considerable amount of research proposed with single-pipeline Convolution Neural Networks (CNN) to accommodate this subject. Although these studies achieved results with high accuracy, they required a large dataset and an implementation on dedicated hardware configuration. This paper presents a novel method with vision-based approach to detect moving vehicles from static surveillance cameras. Moving vehicles are detected and analysed by means of using a Neural Decision Tree accompanied with geometric features to classify vehicles and a Single Shot Detector to handle occlusion when inter-vehicle space between vehicles significantly decreases. Experiments have been conducted on the real-world data to evaluate the performance and accuracy of the proposed method. The results showed that our proposed method achieved a promising detection rate with real-time processing on regular hardware configuration.