四帧间差分和改进混合高斯模型对运动目标的检测

针对运动目标检测过程中已有算法难以同时提高准确性和实时性的问题,提出四帧间差分结合改进的混合高斯模型(Gaussian mixture model,GMM)算法,首先利用四帧间差分对预处理的视频帧差分处理,得到背景区域和运动区域;其次,使用改进的GMM,借助计数器调整高斯模型,提高高斯分量的自适应性,根据单位灰度值确定高斯分量个数,并引入敏感参数改进传统混合高斯模型对学习率的依赖;模型更新时借助计数器确定更新时机;最后,对结果使用形态学处理,提高目标提取的精确度。与已有算法的性能相比,查准率和查全率的调和平均值提高了约44.8%,对GMM算法的改进使得模型训练与检测的计算时间分别缩短至原算法的0.16倍、0.27倍,相比传统的混合高斯模型和文献中的方法,计算时间分别缩短至1/54、1/4、16/25,且对多种场景均能有效适应。

Detection of Moving Targets by Four-frame Difference and Modified Gaussian Mixture Model

Aiming to overcome the issue of being hard to improve the accuracy and real-time at the same time when applying the existing algorithms in the moving target detection process, a four-frame difference combined with modified Gaussian Mixture Model algorithm is proposed. Firstly, differential process of the preprocessed video frame was carried out by four-frame difference to obtain the background and moving area; secondly, the modified Gaussian Mixture Model was used, adjusted by the counter, to improve adaptability and determine the number of Gaussian components based on the unit gray value. And the sensitive parameter was introduced to reduce the dependence of the traditional Gaussian Mixture Model on the learning rate. The counter was also used to determine the update timing; finally, apply morphological processing to the final result to enhance the accuracy of target extraction. Experimental results show that the harmonic mean of the precision and recall is increased by about 44.8% compared with the performance of other algorithms. The improvement of the GMM algorithm shortens the calculation time of model training and detection to 0.16 times and 0.27 times of the original algorithm, respectively. Compared to the traditional Gaussian Mixture Model and method in the literature, the calculating time is shortened to 1/54 ,1/4 and 16/25 respectively. Furthermore, this means can adapt to various scenes effectively.