Window Based Min-Max Feature Extraction for Visual Object Tracking

Abstract

Background/Objectives: Visual object tracking is considered as difficult problem and becomes more challenging because of the various environmental conditions. In order to achieve the efficient result in the area of visual tracking, feature extraction will play the important role. This study demonstrates the Min-max feature extraction method to improve the tracker robustness in the area of visual tracking. Methods: The proposed min-max features along with existing Histogram of Oriented Gradient (HOG), and Convolution Neural Network (CNN) features are given to the Spatial Temporal Regularized Correlation Filter (STRCF) to find the new position of the target and it is successfully solved through Alternative Direction Method of Multipliers (ADMM). By employing both Spatial and temporal regularization methods, without much compromise in the efficiency, the boundary effect is handled. The min-max feature will extract the object’s window-based features as foreground and background. The foreground consists of higher color values than the background. As compared to the Color Names (CN) proposed minmax feature method gives accurate features to identify the objects in a video. In order to present the performance, the method is tested on the OTB dataset image sequences and compared with the state-of-the-art tracker and achieved the promising results for all of tested videos. Findings: Our method, using Minmax feature, gives Mean OP and FPS 61.44% &18.27 respectively, which shows improvement in the tracking accuracy along with the computational speed as compared to CN feature. Keywords: MinMax feature; CNN; Histogram of Oriented Gradient; Occlusion; Visual tracking; Video surveillance