This paper presents a hybrid feature fusion approach for visual object tracking that combines hand-crafted gradient-based descriptors with semantic representations extracted from a pretrained convolutional neural network. The proposed tracker integrates Histogram of Oriented Gradients (HOG) features with ResNet-50-based deep features within a correlation filter framework to improve robustness against appearance variations, occlusion, and illumination changes. A concatenation-based fusion strategy with adaptive confidence-driven weighting is incorporated to dynamically balance the contribution of handcrafted and deep features during tracking. The architecture employs parallel feature extraction branches and multi-scale feature integration to enhance localization performance while maintaining computational efficiency. Experimental evaluation on standard benchmark datasets, including OTB-2015 and related tracking sequences, demonstrates that the proposed fusion strategy provides improved performance compared with individual feature-based tracking approaches and achieves competitive results relative to baseline correlation-filter trackers under challenging conditions. The study also outlines potential directions for further enhancement through online adaptation and attention-based feature fusion strategies.
Sarma et al. (Wed,) studied this question.