高斯混合模型与GhostNet结合的YOLO-G遗留物检测方法

YOLO-G Abandoned Object Detection Method Combined with Gaussian Mixture Model and GhostNet

遗留物检测作为视频监控的关键支撑技术之一,具有广泛的应用前景.针对现有方法存在无法较好地解决光影变化问题、深度学习遗留物检测方法中神经网络的参数量较多及准确度低等问题,提出基于YOLO-G的遗留物检测方法.首先结合高斯混合模型背景建模进行前景检测,根据移动区域与静止区域的分离距离与时间得到可疑静止区域,将判定为分离时刻的帧图像传入深层神经网络进行检测与识别;然后在网络模型中将幽灵网络中的幽灵模块应用于CSPDarknet53主干网络;最后引入压缩激励网络进一步提高特征提取能力.实验结果表明,所提方法的检测准确率比FCOS,SSD,RefineNet,YOLOv3,LRF和YOLOv4分别提高了34.22%,23.86%,16.64%,13.19%,8.16%和1.41%,网络参数量比YOLOv4减少了22.78%.

As one of the keys supporting technologies in video surveillance,recent years have witnessed a wide range of applications in leftover detection.The existing methods are not able to solve the problem of light and shadow changing.Besides,the number of parameters of the neural network in deep learning-based legacy detection rises up and the precision is not enough.This paper proposes the abandoned object detection method based on YOLO-G.The method combines background modeling of Gaussian mixture model for foreground detection.With the separation distance and time between the moving area and the static area,a suspicious static area is obtained.The frame image judged as the separation moment is sent to the deep neural network for further detection and recognition.This network model adopts the ghost module(GhostModule)in the ghost network(GhostNet)and applies it to the CSPDarknet53 backbone network.Finally,the squeeze-and-excitation network layer(SElayer)is introduced to further improve the feature extraction ability.The experimental results show that the detection accuracy of YOLO-G is 34.22%,23.86%,16.64%,13.19%,8.16%and 1.41%higher than that of FCOS,SSD,RefineNet,YOLOv3,LRF and YOLOv4,and the network parameter is 22.78%lower than YOLOv4.