基于图像识别的施工防护栅栏异常检测的研究

Study on protective fence anomaly detection in construction site based on image recognition

在许多施工场景的安全监控中,需要及时发现防护栅栏的异常状态并进行告警。将图像识别技术应用于视频监控系统中,针对栅栏图像特有的特点,设计了基于HOG和基于深度学习的检测方法。通过实验比较分析,基于深度学习的检测方法比基于HOG的检测方法具有更高的检测精度,基于深度学习的检测方法其最高检测精度达到99.8%。合理设计深度神经网络的参数规模和网络架构,即使在无GPU的运算环境下也能达到准实时的检测需求,在无GPU的运算环境下,其对高清图像的检测速度能达到1.7s/f。

In security monitoring construction site,it is necessary to find an anomaly of protective fence and give an alarm in time.Image recognition technology is applied to video surveillance system for intelligence.The two anomaly detection methods based on HOG（Histogram of Oriented Gradient）and based on deep learning are designed according to the characteristic of a fence image.Through experimental comparison,the deep learning based method has higher detection accuracy than the HOG based method.The highest detection accuracy is up to 99.8% based on the deep learning method.Reasonable design of parameter scale and network architecture of deep neural networks can achieve near real time detection requirements even in the absence of GPU computing environment.The detection speed of high definition images can reach 1.7 s/f in the absence of GPU computing environment.