Article Preview
TopIntroduction
Target tracking is an important task in the computer vision. It means the process of constantly extrapolating the target state in video sequences S. Zhang et al. (2013), and its main task is to estimate the target motion trajectory in video sequences. In general, a target tracking algorithm is composed of three basic parts: motion model Widynski, Dubuisson, and Bloch (2011); Isard and Blake (1998), appearance model Mei and Ling (2009); Comaniciu, Ramesh, and Meer (2003), and update strategy Ross et al. (2007); Babenko, Yang, and Belongie (2011). Appearance model is a key to design a robust appearance model. Unfortunately, after many years of research S. Zhang et al. (2013); W. Liu, Chai, and Wen (2013), existing tracking technologies are still faced challenge by some factors, such as occlusion, motion blue, rotation, illumination variation, shape deformation and background clutter. One of the most important reasons is that the tracking algorithm has many difficulties in distinguishing between target and background.
Robust object tracking is merely a basic function of the human visual system (HVS). Studies show that HVS have outstanding ability to data filtering and information locating. Visual attention is a critical mechanism to ensure efficient working for human eyes, which directs the handling of most relevant visual data, in particular, directs our gaze rapidly towards objects of interest. As we know superpixels are just like image blocks which made up of adjacent pixels with similar color and texture features, and superpixels segmentation method can capture redundant information of image in order to reduce the complexity of the task.
The research of target tracking algorithm based on superpixel is insufficient. The main problem is that the load of computation is large and it is easy to drift for the scene with similar object and background. Inspired by the above-mentioned problems, this paper proposes a target tracking algorithm based on superpixel segmentation and visual attention mechanism. The visual saliency mechanism is used to avoid a large number of computations, and it is easier to distinguish the targets in the above scenarios with HOG features. So, the purpose of this paper is to build an efficient appearance model based on superpixel segmentation and visual attention mechanism, using particle filter framework as a motion model for tracking tasks. The whole tracking process shows in Figure 1.
Figure 1. The workflow of our tracker
TopThe computer vision literature V. Mahadevan and Vasconcelos (2009) has postulated a connection between discriminant tracking and one of the core processes of early biological vision – saliency, by suggesting that the ability to track objects is a side-effect of the saliency mechanisms that are known to guide the deployment of attention. More precisely, V. Mahadevan and Vasconcelos (2009) has hypothesized that tracking is a simple consequence of object-based tuning, over time, of the mechanisms used by the attentional system to implement bottom-up saliency. So Mahadevan et al. refer to this as the saliency hypothesis for tracking Vijay Mahadevan and Vasconcelos (2012). If the hypothesis held, these three assertions would be established inevitably:
- 1.
Tracking reliability of saliency target is higher than non-saliency target;
- 2.
Tracking reliability depends on the definition variables of saliency, such as feature contrast;
- 3.
Saliency and tracking will be achieved by utilizing a normal low-level neural mechanism.