Existing object trackers are mostly based on correlation filtering and neural network\nframeworks. Correlation filtering is fast but has poor accuracy. Although a neural network can\nachieve high precision, a large amount of computation increases the tracking time. To address this\nproblem, we utilize a convolutional neural network (CNN) to learn object direction. We propose a\ntarget direction classification network based on CNNs that has a directional shortcut to the tracking\ntarget, unlike the particle filter that randomly finds the target. Our network uses an end-to-end\napproach to determine scale variation that has good robustness to scale variation sequences. In the\npretraining stage, the Visual Object Tracking Challenges (VOT) dataset is used to train the network for\nlearning positive and negative sample classification and direction classification. In the online tracking\nstage, the sliding window operation is performed by using the obtained directional information\nto determine the exact position of the object. The network only calculates a single sample, which\nguarantees a low computational burden. The positive and negative sample redetection strategies can\nsuccessfully ensure that the samples are not lost. The one-pass evaluation (OPE) evaluation results of\nthe object tracking benchmark (OTB) demonstrate that the algorithm is very robust and is also faster\nthan several deep trackers.
Loading....