Map-matching on Wireless Traffic Sensor Data with a Sequence-to-Sequence Model

Map-matching plays an essential role in many location-based applications. It seeks to translate a sequence of timestamped location measurements, which may originate from GPS, Bluetooth, or cellular sources, into the actual routes that moving objects follow on the underlying digital road network. Some existing work focus on map-matching methods based on Hidden Markov Models. While powerful, these methods are computationally demanding and require highly accurate location information. In contrast, neural network-based methods offer the ability to handle more complex data sources, but face challenges when applied to large-scale road networks. In this research, we delve into the task of map-matching using wireless traffic sensor data, specifically Bluetooth, in the context of expansive road networks. We introduce a Turn-Based Map-Matching (TBMM) model, built upon a Sequence-to-Sequence framework. This model accepts a sequence of Bluetooth readings as input and generates a sequence of successive turns with a predicted start road segment. As the sequence of turns is generated, the corresponding route is concurrently reconstructed, adhering to the topological structure of the underlying road network. Furthermore, we employ a two-step training approach to optimize our model. We begin by pre-training the model by minimizing cross-entropy loss. Subsequently, we deploy reinforcement learning to fine-tune the model, thereby further enhancing its performance. Our experimental study shows the promising performance of our TBMM model compared with two state-of-the-art solutions, achieving approximately 98% in precision, recall, and F1-score, demonstrating the potential of our approach in the domain of map-matching.