멀티모달 학습을 활용한 다중 센서 융합

Abstract

In this study, we present a method for producing more accurate results by changing the weights of each sensor according to the surrounding environment, beyond utilizing multiple models and multiple sensors. In multi-sensor fusion, the weight of each sensor varies depending on the accuracy of each sensor. Covariance intersection and modified version widely used because it is useful when one cannot know the exact value of cross covariance. However, they only get fixed weights by initial parameters, making it difficult to cope with environmental changes: climate, road geometry, and driving conditions. This innate limit causes the location error of targeting vehicles. To overcome this, our method learns various information obtained from sensors through multimodal learning and performs data fusion that adapts to various situations in real time. Proposed method effectively recognizes the surrounding environment and shows higher accuracy, as well as real-time computational capabilities. This method has been verified with the actual vehicle data.