Machine Learning-Enabled Automatic Vehicle Detection for Virtual Weigh-in-Motion Applications

Weigh-in-Motion (WIM) techniques are emerged as a prominent solution to regulate weight-related violations, but high installation and recurring maintenance costs are the major obstacles of employing them. Also, the installation process of the inductive loop for vehicle presence detection can cause damage to pavement structure and lead to early deterioration of the pavement surface. Recently, Virtual Weigh-in-Motion (V-WIM) techniques consisting of WIM scales, traffic surveillance videos, and other sensors become a new technological trend that attracted interests from state DOTs to deal with size and weight enforcement that reaches beyond the WIM’s conventional role in data collection. Accurate detection of the specific vehicle-type information to activate the WIM sensors is the first major step for the V-WIM from roadside captured images/videos containing vehicles. However, issues of existing vehicle detection such as the lack of vehicle-type recognition, low-detection accuracy, night and dusk/dawn conditions, and slow speed are the major challenges for developing a powerful V-WIM. This paper applied machine learning based You Only Look Once (YOLO) algorithm trained on a custom dataset as an improved automatic vehicle detection process to work on existing traffic surveillance videos/images as the vehicle sensor data source. The results showed that it could detect vehicles close to the camera with high confidence. Such an improved automatic vehicle detection process may enhance the performance of V-WIM support to develop a dynamic traffic monitoring system and increase the applications of traffic surveillance videos/images.