Analysis of Measurement Errors Influence on the Quantitative and Qualitative Results of Car-Following Model Calibration

In the past two decades there has been a significant increase in using real-traffic data for car-following models calibration. The most widely used way of microscopic trajectory data collection utilizes digital cameras for recording the video of vehicles on the road and subsequent digital image processing to extract vehicle trajectory data. Unfortunately, this method of data collection is not perfect and obtained trajectories contain multiple forms of errors and noise due to both method of data collection and the algorithms used for processing digital data. Such data, however, is widely used in the transportation community for developing and calibrating various models. Some researchers use data post-processing prior to calibration to address the problem noisy while others do not. From this arises a question of measurement errors influence on the quality of results of calibration using real-traffic data. The objective of this study is to compare quantitative results of calibration of two car-following models performed using raw data versus post-processed data, and provide recommendations on post-processing data prior to calibration. Calibration of two car-following models is first performed on the raw data and then on several datasets that filtering techniques had been applied to in order to reduce measurement errors. Two smoothing methods utilized in this study include moving average with a Gaussian kernel and exponential moving average; data frequency reduction is also considered as an alternative. Calibration is performed for Tampere model (stimulus-response model) and Gibbs model (belongs to safe-distance models). The study examines and compares the results of calibration using raw and filtered data; the recommendation is made to perform post-processing of the raw data prior to using the data for calibration procedures