|
The widespread availability of cost-effective sensing technologies is translating into an increasing number of highway bridges being instrumented with structural health monitoring (SHM) systems. Current bridge SHM systems are only capable of measuring bridge responses and lack the ability to directly measure the traffic loads inducing bridge responses. The output-only nature of the monitoring data available often leaves damage detection algorithms ill-posed and incapable of robust detection. Attempting to overcome this challenge, this study leverages state-of-the-art computer vision techniques to establish a means of reliably acquiring load data associated with the trucks inducing bridge responses. Using a cyberenabled highway corridor consisting of cameras, bridge monitoring systems, and weigh-in-motion (WIM) stations, computer vision methods are used to track trucks as they excite bridges and pass WIM stations where their weight parameters are acquired. Convolutional neural network (CNN) methods are used to develop automated vehicle detectors embedded in GPU-enabled cameras along highway corridors to identify and track trucks from real-time traffic video. Detected vehicles are used to trigger the bridge monitoring systems to ensure structural responses are captured when trucks pass. In the study, multiple one-stage object detection CNN architectures have been trained using a customized dataset to identify various types of vehicles captured at multiple locations along a highway corridor. YOLOv3 is selected for its competitive speed and precision in identifying trucks. A customized CNN-based embedding network is trained following a triplet architecture to convert each truck image into a feature vector and the Euclidean distance of two feature vectors is used as a measure of truck similarity for reidentification purposes. The performance of the CNN-based feature extract is proved to be more robust than a hand-crafted method. Reidentification of the same vehicle allows truck weights measured at the WIM station to be associated with measured bridge responses collected by bridge monitoring systems.
|
