Transfer Learning in Data Fusion at Autonomous Driving

one of the major challenges for Autonomous Driving (AD) system while adapting it in various environments at any time is the perception of and prediction of objects and road user behavior and their actions, A special attention is put into data and sensor fusion methods for object detection and identification and scene understanding. In addition, we highlight on the widely used deep learning network architectures in autonomous driving. In this research, we focus on different data fusion methods including early fusion and late fusion methods and how to implement the late data fusion technique from a source domain such as a reference vehicle or a stationary car with well-trained dataset and large amount of labelled data by using transfer learning at late stages of Neural Network on the target domain which has a few labelled data. In fact, the late fusion method is useful in enhancing the training rate for the target domain with assist of the well-trained source domain instead of training target domain from scratch. Also, the data fusion using transfer learning will increase the performance and accuracy level that can be achieved in the target domain while it learns the target task with assistance from the source domain.