A Discriminative Tracking Algorithm Based on Multi-Feature Fusion Mechanism

With the rise of embedded platforms, offline application of visual tracking technology directly on terminal platforms will become a popular remote sensing technology in the future. The current development directions usually focus on improving accuracy. Without the GPU platform, the real-time performance of the algorithm will significantly degrade. We propose a discriminative tracking method based on a multi-feature fusion mechanism, following the basic framework of correlation filter trackers. The innovation lies in two aspects: Firstly, unlike conventional direct feature fusion, we design a feature weighting mechanism to fuse two independently trained features, namely the FHOG feature and the color histogram feature. Experimental results demonstrate that this approach is more effective in exploiting the complementarity between features. Secondly, to reduce unnecessary computations and improve the inference speed of the tracker, we design a peak discrimination strategy based on average peak correlation energy to manage the scale adaptation mechanism and feature model updates. Considering our tracker needs to track targets over long distances in the air, all experiments are conducted using the UAV123 dataset as the test set. The experimental results demonstrate that the proposed model can be deployed on small embedded systems without neural network inference capabilities, ensuring both tracking accuracy and real-time performance.