空间约束混合高斯运动目标检测

目的 针对传统混合高斯模型前景检测运算量过大问题,提出一种基于空间约束的混合高斯前景检测算法。方法 通过快速初始化缩短模型的初始建立过程;采用双重背景模型机制,以自适应背景减法的前景检测结果作为混合高斯前景检测的空间约束条件,降低模型在背景区域的冗余运算;运用多策略自适应模型更新,提高前景检测的准确性。结果 在各种测试场景下,与传统混合高斯法、CodeBook、GMG、偏差均值混合高斯模型(MODGMM)等算法相比,该算法具有更好的准确率以及4倍以上的处理速度。结论 在固定相机场景下的运动目标检测中,算法能有效提高传统混合高斯法的准确性且具有极高的实时性。

Moving object detection using improved Gaussian mixture models based on spatial constraint

Objective Aiming at reducing the high computation cost of the classic Gaussian mixture model (GMM), we propose a GMM with spatial constraint (SCGMM). The main reason for the high computation cost of the GMM is that all pixel models are computed at every frame, and a large part of the computation is useless for GMM. Thus, the SCGMM focuses on reducing the number of models involved in the computation of the GMM. Method For the three parts of the GMM, three methods are utilized to reduce the computation cost of the GMM. In the initial part of the GMM, a method of fast initialization is utilized to shorten the process of initial modeling. The initialization of the GMM requires a large amount of statistical information from all the pixel points. Each pixel should be involved in all operations and the computation cost for one frame used in the initial part cannot be obviously reduced. For this reason, a simple adaptive learning rate is applied to reduce the number of frames required in the initial part of the GMM. In the moving object detection part of the GMM, a double background model is adopted. The detection results for moving objects of the first adaptive background model are used as the spatial constraint condition of the GMM to reduce the redundant computation of the GMM at the region without moving objects. The moving object detection method of the GMM is also used at the region that may contain moving objects to maintain the accuracy of the GMM. Therefore, the advantages of the SCGMM in the moving object detection part is that the SCGMM reduces the number of pixels involved in the computation of the GMM and maintains the accuracy of the GMM. In the parameter update part, multi-strategy adaptive model updating is adopted. The final result of the moving object detection part is used as the spatial constraint condition to reduce the quantity of pixels involved in parameter update. Adaptive learning rate and periodic global update are applied to improve the accuracy of moving object detection. By using the aforementioned methods, the performance of GMM is evidently optimized. Result Experimental results show that SCGMM has better performance and accuracy than GMM, CodeBook, GMG, and MODGMM(mean of deviation GMM). The processing speed is increased by more than three times. Notably, the processing speed of SCGMM is increased by more than six times compared with that of GMM, and the percentage of pixels involved in the complex computation process is less than 20%. Conclusion Compared with GMM, SCGMM has better performance at real-time processing and better accuracy in a fixed camera scene.