Acceleration-Based Deep Learning Method for Vehicle Monitoring

Abstract

An automatic vehicle monitoring system can provide supports not only for intelligent transportation systems, but also for bridge weigh-in-motion (BWIM) systems, which use structural response to identify vehicle weights. In this paper, we provide a vehicle monitoring solution for acceleration-based BWIM system, utilizing deep learning and wavelet transform methods. The monitoring task is divided into three subtasks, including valid sequence detection, valid axle location, and driving lane identification. In first procedure, a shallow convolutional neural network is trained using time-frequency spectrograms to discover valuable time series. After that, an adaptive wavelet transform method is employed to locate axles from each valid sequence. Finally, the driving lane can be determined by cross-comparing vibration responses. Comparing with solutions based on all-in-one deep networks, the proposed method is computationally efficient and has improved generalization capability owing to the three-step division of the task. Evaluation is conducted on a multi-lane highway bridge located in Tokyo. Results show that 97% of vehicles can be identified correctly. For all recognized vehicles, the accuracy of driving lane detection is 100%.