Siamese Networks for RF-Based Vehicle Trajectory Prediction

Traffic surveys monitor the traffic flow to generate data used in improving traffic management. These surveys were traditionally carried out solely by manpower. With the improvement of video analysis, traffic surveys have been shifting towards using automated processes. Surveys carried out by manpower are ineffective unless the surveyors are professionally trained while using video footage requires expensive systems, large digital storage space, and bandwidth for IoT applications. Smart city concepts rely heavily on real-time data from many systems to operate and would benefit from lower computational power, power to operate, and bandwidth requirements. This research investigates the use of radio frequencies (RF) to capture vehicles moving through an intersection. RF data has an inherent nature of being more privacy-preserving than video and requiring less digital storage space and bandwidth while having all the benefits of using video for this purpose. Multiple machine learning models were experimented with for vehicle detection where long short term memory neural networks achieved the best result and was used to detect the presence of vehicles. The vehicle trajectory prediction algorithm used the data of the detected vehicles to predict the trajectory using the similarity between records where siamese neural networks with triplet training outperformed other methods. The data generated can be used to compute metrics for the vehicle occupancy of an intersection. This research enables traffic surveys and real-time monitoring to be carried out with minimal manpower using low-cost low-powered devices that generate smaller sized data samples.