基于深度神经网络的智能监控系统的研究与实现

发布时间：2018-08-23 08:52

【摘要】：如今,社会的安全形势日益复杂,传统的监控识别算法的识别效果已经不能满足需求,随着计算机视觉技术和深度学习的理论和研究的迅猛发展,可以用更加智能的算法替代传统识别算法。本文主要研究了基于深度网络的监控识别算法,并将算法应用在智能监控系统当中。与传统的监控识别算法不同,本文的深度网络不需要人为的对特征进行设计,也不需要根据环境和算法人为的调整参数,而是设计了一个多尺度的卷积神经网络,在行为视频库上训练出了可以提取人体行为的网络模型,将其作为监控异常行为的算法。此外在YOLO网络的基础上,再用行人图片库优化了网络的特征提取能力后,设计了入侵检测算法和徘徊检测算法,算法都在仿真实验中取得了优秀的识别效果,最后将算法都应用在了智能监控识别系统之中。本论文重点实现了基于改进多尺度神经网络的行为识别算法和基于YOLO的入侵检测算法和徘徊检测算法,具体的工作如下:1.研究3D CNN,通过用更长的视频输入,对灰度通道和光流通道进行独立卷积连接下层网络的方式,改进了输入层,使得第一层卷积层的卷积核可以有专一性的特征提取能力;在网络结构中增加了三维时空下采样层,给网络在时间尺度上增加了平移不变性,提升了网络对视频识别的鲁棒性;增加了NIN结构,使网络有非线性特征的提取能力;增加了时空金字塔结构,使得整个网络中可以输入不同分辨率,不同时长的视频,使得本模型适用于真实环境的视频监控,加强了网络的应用价值。2.对YOLO网络进行新的训练,加强网络中的人体特征提取能力,将网络作为特征提取器与传统的入侵检测算法结合,实现了一种新的入侵检测算法,利用这个算法,结合了运动方向特征和颜色特征,实现了一种自动检测人体并且抗遮挡的徘徊检测算法。这种算法与传统的入侵检测相比,本算法可以区分入侵主体是否是人体,并且有更好的识别效果,具备更大的实用价值。3.根据现实的智能监控系统的需求分析,设计了服务端的整体框架,将自己设计实现的基于三维多尺度卷积神经网络的人体行为识别算法和基于YOLO模型的入侵检测算法和徘徊检测算法,应用在了智能监控系统的识别服务端上。并且通过加速优化实现了系统支持16路以上实时识别的能力。
[Abstract]:Nowadays, the security situation of the society is becoming more and more complicated, and the recognition effect of the traditional monitoring and recognition algorithm can not meet the demand. With the rapid development of the theory and research of computer vision technology and depth learning, A more intelligent algorithm can be used to replace the traditional recognition algorithm. This paper mainly studies the monitoring and identification algorithm based on the depth network, and applies the algorithm to the intelligent monitoring system. Different from the traditional monitoring recognition algorithm, the depth network in this paper does not need to design the features artificially, nor need to adjust the parameters artificially according to the environment and algorithm. Instead, a multi-scale convolution neural network is designed. The network model which can extract human behavior is trained on the behavior video database and used as an algorithm to monitor abnormal behavior. In addition, on the basis of YOLO network, the feature extraction ability of the network is optimized by using the pedestrian picture library, then the intrusion detection algorithm and the hovering detection algorithm are designed. The algorithm has achieved excellent recognition effect in the simulation experiment. Finally, the algorithms are applied in the intelligent monitoring and identification system. This paper focuses on the implementation of behavior recognition algorithm based on improved multi-scale neural network, intrusion detection algorithm and hovering detection algorithm based on YOLO. The specific work is as follows: 1. The 3D CNN is studied. By using longer video input to connect the gray channel and the optical flow channel independently to the lower layer network, the input layer is improved, so that the convolution kernel of the first layer can have a unique feature extraction ability. Three dimensional sampling layer is added to the network structure, which increases the translation invariance of the network in time scale, improves the robustness of the network to video recognition, increases the NIN structure, and makes the network have the ability of extracting nonlinear features. The structure of temporal and spatial pyramids is added to make the whole network input video with different resolution and time, which makes this model suitable for the real environment video surveillance, and strengthens the application value of the network. 2. The new training of YOLO network is carried out, and the ability of human body feature extraction in the network is strengthened. A new intrusion detection algorithm is implemented by combining the network as a feature extractor with the traditional intrusion detection algorithm. A hovering detection algorithm which can detect human body automatically and resist occlusion is realized by combining motion direction feature and color feature. Compared with traditional intrusion detection, this algorithm can distinguish whether the intruder is human or not, and has better recognition effect, and has more practical value. According to the requirement analysis of the real intelligent monitoring system, the whole frame of the server is designed. The human behavior recognition algorithm based on 3D multi-scale convolution neural network, the intrusion detection algorithm based on YOLO model and the hovering detection algorithm are applied to the identification service of intelligent monitoring system. And the system can support more than 16 real-time recognition by accelerating optimization.
【学位授予单位】：电子科技大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.41;TP183

【相似文献】