基于DeepMask和RJMCMC的遗留箱体检测

Abundant Box Detection Based on Deep Mask and RJMCMC

论文针对视频监控中的遗留箱体检测进行了研究,提出了基于深度神经网络特征提取及分割和贝叶斯网络建模相结合的检测方案。深度神经网络用于特征提取及个体分割以获取当前帧的似然概率及箱体检测,并使用贝叶斯建模方法将跟踪问题转化为状态的最大后验估计,在求解过程中采用RJMCMC的迭代采样法,以实现对可变多目标的跟踪。进而借助于RJMCMC过程的三种行为方式中的"新生"及跟踪状态,来判别箱体是否为遗留,从而实现对视频中遗留箱体检测。实验结果集定量分析评估表明了该算法的有效性。

In this paper,the detection of legacy box in video surveillance is studied,and a detection scheme based on depth neural network feature extraction and segmentation and Bayesian network modeling is proposed. The depth neural network is used for feature extraction and individual segmentation to obtain the likelihood and box detection of the current frame,and the Bayesianmodeling method is used to transform the tracking problem into the maximum posteriori estimation of the state,in the process of solv-ing the use of RJMCMC iterative sampling side,in order to achieve a variable multi-target tracking. And then by means of the RJM-CMC process of the three kinds of behavior in the"new"and tracking status,to determine whether the box is left,so as to achieve the video in the box detection. The quantitative analysis of the experimental results shows that the algorithm is effective.