Road intersection identification from crowdsourced big trace data using Mask-RCNN

Abstract

Road intersection data are critical in many spatial applications and analyses. Approaches to identifying road intersections from various sensor data have been widely discussed in many existing studies, which focused on macroscopic information detection (e.g., location and size confirmation) and microscopic structure extraction (e.g., traffic rules at lane level). As the premise and basis for microscopic structure construction, the accuracy of macroscopic information has an important influence on extracting the detailed spatial structure of road intersections. In this article, we proposed applying a mask region convolutional neural network (Mask-RCNN) framework to automatically detect the macroscopic information of road intersections from crowdsourced big trace data. There are two key points: (1) Mask-RCNN-based road intersection detection; and (2) result optimization and localization. Two real-world GNSS (global navigation satellite system) trace datasets collected in Wuhan and Rome, respectively, were used to verify the applied Mask-RCNN system. The results showed that the identification of various common types of road intersections achieved an overall precision of 97, 99, and 96%; recall of 93, 87, and 90%; and F1 score of 95, 92, and 93 in Wuhan, Shanghai, and Rome, respectively. These indices revealed a better model performance than the existing popular RCNN-based models.