传统Mean-Shift跟踪算法缺少核函数带宽更新策略,故无法解决无人艇跟踪的水面运动目标轮廓变化各向异性问题,提出一种各向异性带宽自适应的Mean-Shift跟踪算法.先用黎曼积分将特征子模型概率密度的归一化常数.近似为积分形式,从而获得不同尺度参数Ch对应的间关系式.然后用梯度上升法使目标模型和目标候选模型之间的相似度函数达到局部最大,由此估计目标在下一帧的带宽与位置.最后为防止带宽更新时结果过小或过大,引入两个正则化参数修正尺度参数.实验结果表明,所提算法对外形轮廓非同比变化的水面运动目标跟踪具有各向异性的带宽自适应调节能力,型心位置准确率较传统Mean-Shift和各向同性带宽自适应Mean-Shift提高了约77. 2％和31. 1％,运行速度可达 20. 7 fps,显示了其鲁棒性和实时性.%The traditional mean-shift (MS) tracking algorithm lacks an effective bandwidth update strategy for surface moving targets in unmanned boats with anisotropic changes in contour. To address the challenge, we propose an anisotropic bandwidth-adaptive MS tracking algorithm. Firstly, the Riemann integral is used to approximate the normalized constant Ch of the probability density of the feature sub model into the integral form. Thus, the relationship between Ch corresponding to different scale parameters h is obtained. Secondly, the gradient ascent method is applied to make the similarity function between target model and target candidate model reach the local maximum, thereby the bandwidth and position of the target in the next frame can be estimated. Finally, two regularization parameters are introduced to correct the scale parameters to prevent the bandwidth updating from being too small or too large. Experimentsal results show that compared with the traditional MS and isotropic bandwidth-adaptive MS, the accuracy of the center location of our algorithm is increased by about 77. 2％ and 31. 1 ％, and its running speed is up to 20. 7 fps,which proves its ability of anisotropic bandwidth-adaptive adjustment, as well as its robustness and real-time.
展开▼
