Multilayer deep feature integration framework for discriminative correlation filter tracking

Due to the complexity of deep features, discriminative correlation filter (DCF)-based trackers only extract single-layer deep features for object description and the characteristic of different layers cannot be fully exploited. Therefore, a framework of integrating multi-layer deep features is proposed for trackers based on DCF. In the training stage, multi-layer deep features are extracted using convolutional neural network. Max pooling and channel compression is applied to deep features to reduce feature dimensions and channels. Then the compressed deep features are cascaded with hand crafted features for correlation filter training. In the tracking stage, only hand-crafted features are extracted in multi-scales and deep features are extracted in single scale. Then tracking responses from single scale deep features are added to responses from multi-scale hand crafted features. Experiment results show that the tracking precision and success rate is 3.7% and 3.9% higher than original LADCF respectively. Meanwhile, the speed is 48.6% faster. It indicates that the proposed method is feasible which can maintain the performance of DCF trackers and improve the speed of the algorithm.