Target re-aware deep tracking based on correlation filters updated online

Deep trackers often use convolutional neural networks (CNNs) pre-trained to extract features. The training dataset always does not contain the tracking objects. Even the objects appearing in the training dataset may always be in arbitrary forms. Therefore, the pre-trained convolutional neural networks extract features with less effectiveness in describing the tracking object. Target-aware deep tracking (TADT) algorithm proposes a scheme to acquire target-aware deep features by an improved regression loss and a ranking loss. Target awareness is achieved by calculating the back-propagation gradients at each pixel in the regression. Multi-channel deep features gradients captured in the first frame affect the efficiency of the tracking in subsequent frames. If the target-aware scheme is updated online, the tracking efficiency can be further improved. In this work, we present a target-aware scheme updated online to re-select deep features generalizing the object appearance more effectively. The base deep features are extracted by the pre-trained VGG16 and further processed to acquire the target awareness deep features. The awareness deep features are re-selected as re-awareness deep features in the framework of correlation filters with parameters updated online. Correlation filters updated online replace improved regression loss to re-identify the importance of features by global average pooling deep features weights. The re-awareness deep features are integrated with a Siamese matching network to determine the target's location and scale in the subsequent frame. Experimental results demonstrate the effectiveness of our presented algorithm compared with TADT in terms of accuracy and speed.