A short-term vehicle speed prediction method based on Kalman filtering and LSTM deep learning

In high-noise scenarios, short-term vehicle speed prediction is easily affected by complex driving behaviors and sudden traffic incidents, resulting in prediction lag or overfitting, and it is difficult to meet the real-time requirements. To this end, a short-term vehicle speed prediction method based on Kalman filtering and Long Short-Term Memory (LSTM) deep learning is proposed. Analyze the influencing factors of vehicle speed from the force state of the vehicle, determine the collected variables, use unmanned aerial vehicles to collect the full-time and spatiotemporal data of the vehicle, decompose the original vehicle speed data with the Variational Mode Decomposition (VMD) module, expand the subsequence into a three-dimensional data tensor according to the sliding window, and improve the DeepSort algorithm to achieve short-term vehicle speed prediction. The experimental results show that the absolute error of the proposed method is 4.26 km/h, and the relative error is 15.8%. The predicted vehicle speed in each interval is closer to the actual vehicle speed, which has important theoretical value for improving the safety of autonomous driving decisions and optimizing traffic signal control.