Researchon Video Segmentation under Occlusion
|Course||Signal and Information Processing|
|Keywords||Video shaking Partial occlusion Corner detection Mean-Shift tracking Particle filter|
There has been a rapid increase in the need of accurate, reliable and real-time object tracking algorithm in recent years. Algorithm of occlusion tracking based center weighted and Bayesian recursive have become important technology for object tracking because of their higher precision and smaller amount of computation. However, according to fixed camera, the typical algorithm of Mean-Shift based center weighted does not apply to mobile environment. While algorithm of particle filter based Bayesian filter theory for ordinary computers is still time-consuming when applied to mobile devices. Therefore, problem mention above is discussed after an overview of previous research. The main work as follows:1. Focuses on the algorithm of moving object extraction based image scale transformation and Gaussian mixture model for camera erratic shaking. As image details can be reduced by image scale transformation and some noise are allowed to exist by Gaussian mixture model, the slight jitter of background can be ignored, and related image processing are used to restore image details.2. The classic algorithm of Mean-Shift is firstly discussed. For Mean-Shift is not robust with target dimensional changing, a method of target probability representation is put forward to solve the problem. According to occlusion, Harris corner detection algorithm is combined to track human target which is partially occluded. As experimental results shown, the optimal position of human can be voted effectively by weighing the tracking results of human main color features with results of sub-block tracking.3. The classic tracking algorithm of particle filter has been explored. Firstly, two state equations are used to represent one-step transition probability for particle filter, and experiments show this method has more accurate than traditional tracking method while number of particles is small; Secondly, diffusing particles’ lookup zones by adding some noise component in the process of particle resampling makes it convergence to target area quickly under condition of partial occlusion; Finally, useless particles are reduced while small number of particles can guarantee the algorithm has a good real-time, experiments show that the method of deleting particle number commonly has a balance point between real-time with tracking precision.