An Ensemble Learning Approach to Improve Tracking Accuracy of Multi Sensor Fusion

Anoop Karnik Dasika, Praveen Paruchuri

Abstract: Finding or tracking the location of an object accurately is a key problem in defence applications as well as the problem of object localization in the fields of robotics and computer vision. Radars fall into the spectrum of high-end defence sensors or systems on which the security and surveillance of the entire world depends. There is much focus on the topic of Multi Sensor Fusion (MSF) in recent years with radars as the sensors. In this paper, we focus on the problem of asynchronous observation of data which can reduce the tracking accuracy of an MSF system comprised of radars of different types at different locations. Our solution utilizes a machine learning approach to train models on hundreds of hours of (anonymized real) Multi Sensor Fusion data provided by radars performing tracking activity across the Indian airspace. Our approach comprises of 3 steps: In step 1, we train an ensemble model of logistic regression and Xgboost to predict Splitting error. In step 2, we use Xgboost to predict the second type of error, namely Merging error to improve the tracking accuracy further. The third step uses nearest neighbour search to compensate for the predicted errors by retaining the data points removed in the first step while maintaining the tracking accuracy. Our experimental results show that the trained models provide reasonable predictions of errors and increase the accuracy of tracking by 15% while retaining 100% of the data.