基于深度帧差卷积神经网络的运动目标检测方法研究

Research of Moving Object Detection Based on Deep Frame Difference Convolution Neural Network

复杂场景中的运动目标检测是计算机视觉领域的重要问题,其检测准确度仍然是一大挑战.本文提出并设计了一种用于复杂场景中运动目标检测的深度帧差卷积神经网络(Deep Difference Convolutional Neural Network,DFDCNN).DFDCNN由DifferenceNet和AppearanceNet组成,不需要后处理就可以预测分割前景像素.DifferenceNet具有孪生Encoder-Decoder结构,用于学习两个连续帧之间的变化,从输入(t帧和t+1帧)中获取时序信息;AppearanceNet用于从输入(t帧)中提取空间信息,并与时序信息融合;同时,通过多尺度特征图融合和逐步上采样来保留多尺度空间信息,以提高网络对小目标的敏感性.在公开标准数据集CDnet2014和I2R上的实验结果表明:DFDCNN不仅在动态背景、光照变化和阴影存在的复杂场景中具有更好的检测性能,而且在小目标存在的场景中也具有较好的检测效果.

Moving object detection in complex scenes is an important problem in computer vision domain,and the detection accuracy is still a great challenge.In this paper,we propose and design a deep frame difference convolution neural network(DFDCNN)for moving object detection in complex scenes.DFDCNN consists of DifferenceNet and AppearanceNet,which can predict and segment the foreground pixels simultaneously without post-processing.DifferenceNet has Siamese Encoder-Decoder structure,which is used to learn changes between two consecutive frames and to obtain temporal information from inputs,while AppearanceNet is used to extract spatial information from the input frame,and fuse the temporal information and spatial information by fusion of feature maps.Finally,multi-scale spatial information is retained through multi-scale feature map fusion and stepwise up-sampling to improve the sensitivity to small objects.Experiments on two public standard datasets:CDnet2014 and I2R demonstrate that the proposed DFDCNN outperforms the classic algorithms significantly from both qualitative and quantitative aspects.The experimental results illustrate that the proposed DFDCNN shows much better detection performance in complex scenes where dynamic background,illumination variation and shadow exist,and there is improvement for scenes,in which small objects exist.