DSP视频监控中运动目标检测方法研究

Moving Object Detection Techniques in Intelligent Video Surveillance System Implemented by DSP

对高斯模型的更新做出改进,以找到一种快速、有效的、适合DSP运算的算法。高斯模型的参数对背景模型判别产生的影响是不同的,因此它们的更新对于背景模型判别产生的影响也是不相同的。学习率直接影响模型参数的更新,如果对所有参数采用同一学习率,当学习率取值比较大,适应环境变化能力强,但容易受噪声影响,不够稳定;当学习率取的比较小,适应环境变化的能力就低,但是具有鲁棒性。利用均值和方差对背景模型判别影响不同这一特性,对均值和方差的更新采用不同的学习率,在保持可行的算法复杂度的情况下,使背景模型能够适应背景的变化。

Improving the update of Gaussian model is a viable way to find a fast and effective algorithm adapted to DSP implementation. Different parameters have different influences on the Gaussian model, so do their updating. And, the updating of parameters is affected by the updating frequency directly. If all of the parameters use the same updating frequency, Gaussian model is more adaptable to the variational background when the value of the updating frequency is a bigger one, but it is unstable and sensitive to the noise; whereas, when the value of the updating frequency is a smaller one, the model is unadaptable to the variational background, but is of robustness. Mean and mean square error have different effects on the distinguishing of Gaussian model, which can be used to the improve-ment. Using different frequencies for mean and mean square error, the model can be more adaptable to the variational background, while keeping the acceptable complexity of algorithm.