Research And System Implementation Of Structured Analysis In Surveillance Video

Currently,video data transmitted over the internet account for about 70%of network traffic.Accordingly,it is no exaggeration to say that video data is the largest data in the big data era.It is not a simple problem for the computer to realize and understand the world better,which has been a hot topic in the field of computer vision in recent years.With the improvement of GPU computing power,the algorithm of deep learning has also been developed rapidly.In most cases,the computer vision method based on deep learning can solve the related tasks of computer vision better and it has become the consensus of industry and academia.In order to make the computer understand the scene better in the camera,this thesis explores core algorithms for the application of computer vision and develops the software,hardware and related algorithms which is driven by social needs and practical projects designs.This thesis is divided into two parts:the first part is the research and development of multi-object detection,multi-object tracking and so on.The second part is the research and development of the three interrelated projects which are excavated through the actual needs of the society.The following problems in computer vision were studied in the completed project:1.Multi-object detection for each frame using the algorithms of deep learning has been achieved in real time.However,object detection for each frame cannot determine the relationship of the same object in the consecutive frames.Thus it is hard to know target motion trajectory in the scene,and the position and the time of appearance disappearance of the object etc.In order to solve this problem,we design a high-speed object tracking algorithm which integrates appearance and motion characteristics and combine the algorithm with multiple target detection algorithm in this thesis.The experiment results show that compared with the traditional motion detection and multi-object tracking algorithms,the proposed method has the advantages of fast detection and satisfied tracking performance.2.The performance of object detection algorithm based on deep learning is well for the larger objects,such as the pedestrian or vehicle detection.But the performance is inferior for the small object in the video,such as the clothes of pedestrians and the vehicle's annual detection marks.Finally,we analyzed that it is difficult to locate the object accurately because the feature accuracy of is not enough during the feature extraction using convolutional neural network.In order to solve this problem,the method of cascade neural network is used to analyze the object details.First,we take the object detected by the first level detection network out of the image,and then put it into the secondary network for refinement identification.The experiment shows that this method solves the problem of identifying objects details in the video.3.In the practical project application,the effect of cascade neural network on fine recognition of distant small objects is still not satisfactory.It is mainly due to the extremely low resolution of distant objects in the camera,as well as the poor performance of deep learning in the detection of small object and the difficulty in identifying the detail features of the object.In order to solve this problem,we design a special hardware system.The system uses the gun camera controlled by the computer to capture the object and track it.Then this system uses the zoom camera to capture those objects one by one and recognize the face.Futhermore in order to solve the problem that the face captured in the outdoor is easily affected by illumination,this thesis introduces GAN network to repair illumination.After comparison,the method in this thesis is superior to the traditional image enhancement algorithm.4.First of all,we design and implement a camera AI high point three-dimensional cloud control system based on augmented reality,so that the system can solve the problem that the traditional monitor and control system is difficult to see widely and clearly at the system level in this thesis.We introduce the core algorithm-an augmented reality coordinate algorithm based on neural network.Secondly,in order to enhance the positioning accuracy of the augmented reality label,we need to know the deep information in the scene.Meanwhile we do not want to increase the hardware cost of the system.In this thesis,we propose a method of monocular deep estimation based on GAN and design a cycle loss function.This method implemented on NYU-Depth data sets has obtained better experimental results and has greater space for improvement compared with the previous method.5.The high altitude parabolic detection system is a part of the community security surveillance system.However,there is no truly effective way to detect high altitude parabolic by video.Although some products based on computer vision use motion detection algorithms to detect parabolic process in the high altitude at present,the actual effect of these products is still not ideal.This is mainly due to the complexity of background change in the surveillance images.So it's difficult to solve the high altitude parabolic detection problems only by the method of computer vision.In this thesis,we introduce a method of detecting the falling time of the object using the Doppler effect of wireless signal echo,and recording the residential building by using a multi-point gun camera with cross arrangement.Then this method calculates the parabolic floor and window to obtain evidence using the binocular parallax method.The experiment result shows that the method of wireless signal processing proposed in this thesis solves the difficulties in computer vision and achieve a good practical application results.