Xiying Wang, Xiwen Zhang, and Guozhong Dai, “Tracking of deformable human hand in real time as continuous input for gesture-based interaction,” in Proceedings of the 12th international conference on Intelligent user interfaces, IUI ’07 (New York, NY, USA: ACM, 2007), 235–242

1. Gesture such as hand gesture input is a natural and effective interactive model

2. However, human hand is articulated, and thus it’s difficult to capture its nonrigid motion due to its continuously changing shape and appearance.

3. Glove-based method use data glove and electronic tracker to capture hand’s
movement, but at the same time, it brings much restriction upon human and makes people feel uncomfortable.

4. Camera-based method uses camera to capture hand motion without any physical restriction upon the user. Compared with gloved-based method, it is a more natural and effective way for gesture-based human-computer interaction.

5. Up to now， visual gesture has already been used in many scenarios, such as virtual mouse for ubiquitous computing environment[1] and remote controller of mobile robot in real world[19].

6. In some of previous work, hand posture is presumed constant [3,21,12] or changes of finger configuration are neglected [14,11]. Others use 3D hand models [1,15] and/or two cameras [10] to track deformable hand gesture, and consequently, the computational complexity of tracking system increases sharply.

7. Existing methods for tracking articulated non-rigid object could be categorized into two groups:
1. model-based methods
2. appearance-based methods [8,7].
Appearance-based tracking can also be divided into region-based tracking,
deformable template or snake model based tracking and so on.

8. As for tracking auto-initialization, it is very important for self-starting, resumption of tracking and further continuous interaction between human and computer. However, most previous work took it for granted or accomplished it manually.

9. Generally speaking, hand tracking for real-time interaction should meet the following three requirements:

a) Tracking speed should be on the real-time level, in other words, high
dimensional feature vector and complex computation should be avoided.
b) Tracking method should be robust enough to be invariant to the transition,
rotation and scaling of tracking target.
c) Tracking should be a continuous process with a auto-initialization. Once
tracking fails, it can be resumed automatically without manual help.

p/s: a lot more interesting info includes the tracking framework, hand track methodology, and fingertip tracking. Might be one of my anchor paper.