动态背景下自适应LOBSTER算法的前景检测

发布时间：2018-08-28 06:43

【摘要】：目的前景检测是视频监控领域的研究重点之一。LOBSTER(local binary similarity segmenter)算法把Vi Be(visual background extractor)算法和LBSP(local binary similarity patterns)特征结合起来,在一般场景下取的了优良的检测性能,但是LOBSTER算法在动态背景下适应性差、检测噪声多。针对上述问题,提出一种改进的LOBSTER算法。方法在模型初始化阶段,计算各像素的LBSP特征值,并分别把像素的灰度值和LBSP特征值添加到各像素的颜色背景模型与LBSP背景模型中,增强了背景模型的描述能力;在像素分类阶段,根据背景复杂度自适应调整每个像素在颜色背景模型和LBSP背景模型中的分类阈值,降低了前景中的噪声;在模型更新阶段,根据背景复杂度自适应调整每个像素背景模型的更新策略,提高背景模型对动态背景的适应能力。结果本文算法与Vi Be算法和LOBSTER算法进行了对比实验,本文算法的前景图像比Vi Be算法和LOBSTER算法的噪声点大幅较低,本文算法的PCC指标在不同视频库中比Vi Be算法提高0.736%7.56%,比LOBSTER算法提高0.77%12.47%,FPR指标不到Vi Be算法和LOBSTER算法的1%。结论实验仿真结果表明,在动态背景的场景下,本文算法比Vi Be算法和LOBSTER算法检测到的噪声少,具有较高的准确率和鲁棒性。
[Abstract]:Objective foreground detection is one of the key points in the field of video surveillance. The LOBSTER (local binary similarity segmenter) algorithm combines the Vi Be (visual background extractor) algorithm with the LBSP (local binary similarity patterns) feature, and it has excellent detection performance in the general scene. However, the LOBSTER algorithm has poor adaptability in dynamic background and more detection noise. To solve the above problems, an improved LOBSTER algorithm is proposed. Methods in the phase of model initialization, the LBSP eigenvalues of each pixel were calculated, and the gray value and LBSP eigenvalue of each pixel were added to the color background model and LBSP background model of each pixel respectively, which enhanced the description ability of the background model. In the pixel classification phase, the threshold of each pixel in color background model and LBSP background model is adjusted adaptively according to the background complexity to reduce the noise in the foreground. The updating strategy of each pixel background model is adaptively adjusted according to the background complexity to improve the adaptability of the background model to the dynamic background. Results compared with Vi Be algorithm and LOBSTER algorithm, the foreground image of this algorithm is much lower than that of Vi Be algorithm and LOBSTER algorithm. In this paper, the PCC index of this algorithm is 0.7367.56 higher than that of Vi Be algorithm in different video libraries, and 0.77% higher than that of LOBSTER algorithm. The PCC index of this algorithm is less than that of Vi Be algorithm and LOBSTER algorithm. Conclusion the simulation results show that the proposed algorithm can detect less noise than Vi Be algorithm and LOBSTER algorithm in dynamic background, and has higher accuracy and robustness.
【作者单位】：江南大学物联网工程学院;
【分类号】：TP391.41;TN948.6

【参考文献】