| With the rapid development of computer technology and information technology, in order to improve the deficiencies of human visual, enabling the computer to process visual information becomes a very remarkable research topic, which promotes the emergence and development of computer vision. Target tracking is an important research area of computer vision, which has been widely applied in intelligent video surveillance, intelligent transportation, human-computer interaction technology, three-dimensional reconstruction, video compression, etc.Most existing tracking algorithms can be described as an optimization process using deterministic or stochastic methods. Deterministic methods aim to searching the minimum of the cost functions using gradient descent methods, such as, the mean shift algorithm. In general, deterministic methods have small computational complexity, but the minimum they found is more likely to be the local extreme value rather than global extreme value. Compared with the deterministic methods, stochastic methods are more robust but have higher computational complexity. Particle filter algorithm is a typical representative of stochastic methods. The key idea of particle filter is to approximate the target state by a set of weighted samples. The state of each particle presents a hypothesis of target state, and the weight presents the probability of the particle state being the true state of the target. The optimal estimate of target state can be expressed as the weighted average of the states of all particles.This thesis analyzes the problems in the particle filter algorithm based on color histogram and puts forward some improvement measures. In the particle filter algorithm based on color histogram, the weight of each particle is updated only by the color feature of the object, which may easily lead to error tracking when the object and the background have similar color distribution or the object is occluded. Scale invariant features are highly discriminative, but they can not describe small targets very well. SIFT features and color features are introduced into particle filter algorithm in this thesis to handle with these two situations. In order to avoid error updating of the color model, whether the color model is updated or not depends on the number of matching feature points between the tracking result in the current frame and SIFT model in this thesis. Experimental results show that the proposed method can effectively improve the tracking precision especially when the object is occluded or in the clutter background of similar colors.The transition model is taken as the importance sampling function in particle filter algorithm, this is not the optimal choice. When the object has rapid and arbitrary motion, the prediction particles drawn from transition model may lie in the tail of the observation model, thus the weights of most particles are low, which leads to the tracking failure. In order to solve this problem, particle swarm optimization algorithm is introduced. Through the PSO iterations, the particles are moved towards the region where the probability of observation model has larger values, thus each particle has larger weight, which results in a better tracking result. Because the PSO algorithm is an iterative process, if the prediction particles are optimized in each frame using PSO, the time complexity of the algorithm will increase greatly. In this thesis, only when the particles’average weight is less than a certain threshold, which means that the states of particles are less likely to be the real state of the target, at this time, PSO algorithm is adopted to optimize the prediction particles and the global extreme value is taken as the estimation of target state when the convergence criterion is satisfied. If the particles’ average weight is larger than the threshold, the real target state can be estimated directly by the prediction particles. Experimental results show that improved algorithm both has robustness and real-time performance. |
PDF Full Text Request