Feature Fusion with Deep Neural Network in Kernelized Correlation Filters Tracker

Abstract

Moving object tracking is the most important component in many computer vision applications. Currently, the ability of computer vision is almost like human vision. Humans can see and track moving objects by looking at notable features such as color, shape, and function. The computer can track moving objects by calculating the characteristics, such as the Histogram of Oriented Gradient (HOG) and grayscale features. These features were used as input in the tracker algorithm. The correlation filter algorithm is extensively used in object tracking applications because of its accuracy and speed. Kernelized Correlation Filters (KCF) is a method that uses correlation for object tracking. The feature fusion is widely used to make tracking more robust. In this paper, the HOG and grayscale features were implemented in the KCF method. Deep Neural Network (DNN) regression was used as a decision feature fusion. With almost similar principle as Non-Maximum Suppression (NMS), where two candidates are detected from overlapping HOG and grayscale features, the region-of-interest (ROI) will be pruned by replacing one ROI to produce a more accurate object candidate. In this study, three TB dataset videos were used for testing, and two videos were used for training. The DNN Regression architecture uses six hidden layers with 512, 256,64,32,16, and 8 nodes. The training accuracy results were 95.76%, with MSE of 9.94 and a loss of 9.93. This research shows that the system can track objects more precisely and robustly with RMSE of 9.38 while achieving 32 FPS.