A Deep Learning Approach for Robust Target Tracking in a Cluttered Environment

In a challenging environment with severe measurement errors, one of the most important factor to improve target tracking accuracy is to understand the channel characteristics of the environment and to estimate the target position in data-oriented manner. In this paper, we propose a deep learning model to estimate target position by processing time difference of arrival (TDOA) measurements in an end-to-end manner, especially for cluttered environments. The proposed TDOA image based target tracking (TITT) model converts each TDOA measurement into TDOA image, then a mask is created to handle TDOA errors. Then, by accumulating and treating TDOA images with masks as input data, we train a CNN model with fully connected layer to estimate target position through back propagation. Simulation results demonstrate that our proposed TITT model outperforms the simple deep learning model in challenging environment with many clutters.