Evaluating the Performance of Deep Learning in Segmenting Google Street View Imagery for Transportation Infrastructure Condition Assessment

Understanding the relationships between the condition of transportation infrastructure and the well-being of citizens in society is of significant importance towards restoring the aging and deteriorating transportation infrastructure in a way that enhances well-being. However, attaining such understanding is challenging because it relies on large-scale transportation infrastructure condition assessment. The broad spatial coverage of Google Street View (GSV) imagery offers a unique opportunity for such large-scale assessment. However, despite the richness of deep learning methods that can segment and recognize transportation infrastructure assets from GSV imagery for subsequent condition assessment, the performance of these methods typically varies. As such, this paper focuses on conducting performance evaluation of representative deep learning-based image segmentation methods to identify the optimal methods for recognizing transportation assets from GSV imagery. The preliminary evaluation results show that the ResNet + UNet and MobileNet + UNet methods achieved the highest intersection over union (IOU) of 0.87.