| Video surveillance is an important means of prison security prevention,and plays an important role in preventing prisoners from escaping and suicide.However,with the increase of video surveillance points,the number of surveillance videos is much larger than the number of monitoring screens displayed in the command center.At present,the average number of video monitoring points per prison is 3,000,while the number of Control Center monitoring screens can be displayed at the same time is about 20.It is difficult for supervisors to achieve effective supervision because of heavy workload and low efficiency when they inspect video through manual mode.With the development of computer vision and artificial intelligence,intelligent monitoring system reduces the work intensity of supervisors and improves the work efficiency of prison management to a certain extent.However,there are several following issues that still can not be solved.First,the image quality assessment system of binocular stereo camera in video surveillance is not fully functional.The quality of monitoring video directly affects the intelligent video analysis in the following stage.The existing image quality evaluation system in video surveillance lacks the image quality assessment methods for face recognition,wide-angle and wide-range binocular stereo camera applied in the important locations.Second,the recognition accuracy of abnormal behavior in video surveillance is relatively low.The intelligent analysis function of existing video surveillance system in the prison can not accurately identify the personnel,and can not automatically alarm the behavior of the prisoners leaving the custody and the police’s irregular duty,so the ability and effect of video surveillance are not really brought into effect.Third,the problem of information loss is caused by the inspection mode of mechanical patrol for video surveillance.The video surveillance in the prison adopts the mode of grouping video location in advance,which is limited by the viewing angle of human eyes and the processing ability of brain,and the probability of detecting abnormal behaviors of prisoners in video surveillance is extremely low.To solve the above-mentioned problems,the thesis starts from the actual work needs and the prison video surveillance system intelligent key technologies,the pain-point problem of the intelligent video surveillance system in the prison,and recommends the importance ranking of the camera locations.This thesis aims to complete the image quality evaluation function of binocular stereo camera,improve the automatic diagnosis ability for image qualityof monitoring video,and overcome the shortcoming of image quality analysis for video surveillance in the prison.Under the condition of ensuring the image quality of monitoring video,the fusion tracking of detachment detection is carried out based on accurate identification of prisoners and police,so as to improve the ability of abnormal behavior analysis for video surveillance in the prison.The main works of this thesis are as follows:（1）To address the problem that the binocular stereoscopic images can not be diagnosed,a no-reference stereoscopic image quality assessment method is proposed for video surveillance.In the thesis,a quality prediction model for distorted stereoscopic images is developed,and the architecture of quality prediction scheme for stereoscopic images is described.For an original stereo pair,the fusion and difference maps are first generated,and the binocular statistical features are extracted as the basic feature vectors.Then,according to the local amplitude and local phase of the stereo pair,the binocular energy responses are calculated as the quality-aware features.Finally,these features of the distorted stereo pairs are mapped to the human perceptual quality scores using the Extreme Learning Machine（ELM）method.Three publicly available benchmark of 3D image databases with subject rating are used as the standard:Live 3D IQ A Phase Ⅰdatabase,Live 3D IQ A Phase Ⅱ database,and MCL-3D data database.The effectiveness and robustness of the proposed scheme,as well as the specific distortion performance on each individual distortion type in the mixed distortion database,are verified by experiments compared with seven representative and recent techniques.Then three different comparison schemes are designed to compare the contribution of each component of the proposed metric to the overall quality score.Finally,the generality and stability of the proposed scheme are verified,and the performance evaluation experiments across data sets are carried out.Experimental results on the benchmark databases show that the proposed model can generate image quality predictions that are well correlated with human visual perception,and it has very competitive performance with the recent relevant techniques.The proposed visual quality prediction is highly correlated with subjective quality judgments for various distortion types of image pairs.（2）In order to solve the problem of high false alarm rate in intelligent video surveillance system,a target analysis method of video surveillance based on iterative learning is proposed.Through iterative learning,training samples are automatically generated for the training process of deep learning model so as to avoid manual labeling.In addition,a detachment detection method based on fusion tracking（Font）is proposed.The target tracking is used to solve the problem of low accuracy in single target detection based on the frame by frame method.In the thesis,the target tracking framework SiamFC based on full convolution with the best performance of deep learning,and the fastest kernel correlation filtering（KCF）method among the classical tracking methods are adopted for fusion tracking.The typical target template is obtained by dimension reduction and clustering of manifold learning method,and the target hash dictionary is constructed by using the perceptual hash feature for online tracking.Two collaborative links are designed,including offline learning and online tracking.The online tracking is used to automatically generate a large number of annotated samples for offline learning iteration,to solve the problem of deep learning method’s demand for annotated data.The offline learning link is used to continuously improve the tracking performance and provide feedback to the online tracking,so that the continuous learning for different scenes can be accomplished.Contrastive experiments show that the detection methods based on the traditional single-frame image target detection mechanism all have many problems,such as false alarm,wrong detection,etc.,but the proposed detection system based on target fusion tracking can effectively solve these problems.（3）To solve the problem of information loss caused by the inspection mode of mechanical patrol in video surveillance,the thesis proposes a display optimization method for video surveillance array based on global-local joint coding model.A global encoder is used to summarize the entire sequence behavior,while a local encoder adaptively selects the important items in the current session,capturing the user’s primary purpose.The representation of sequential behavior can provide useful information to capture the primary purpose of the user in the current session.Therefore,we use the representation of the sequential behavior and the previous hidden state to calculate the attention weight for each click.The characteristics of sequential behavior and user’s purpose are linked together to form an extended representation of each time-stamp.A large number of serialized camera clicks and behavior log samples automatically generated by the monitoring center are used for offline iterative learning,to solve the problem of deep learning method’s demand for annotated data.The off-line learning is used to continuously improve the accuracy of patrol and provide feedback to the online patrol,so as to realize the camera automatic patrol scheme with high precision.By collecting many operation logs of monitoring system in the prison,cleaning and analyzing the log data,it is found that compared with the traditional patrol model,the ranking model of video surveillance proposed in this thesis has better prediction for the patrol data of monitoring camera.Meanwhile,the prediction effect of this model for the patrol data in the daytime is better than that of the patrol data at night.（4）In this thesis,combined with the practical work,the intelligent video analysis platform in the prison is designed and implemented.The platform includes the binocular stereoscopic image quality assessment and the fusion target tracking of abnormal behaviors warning,the intelligent ranking of camera recommendation display,It has been applied in a certain prison and achieved good results,verifiedthe validity of these three video surveillance key technologies.This platform has achieved greater improvement than the existing video surveillance system in the prison in terms of intelligence,the odds of finding abnormal behavior increased by an average of 58.5 percent,and it can also be applied on a large scale in the fields of video surveillance such as social governance and city management,with good social and economic benefits. |
PDF Full Text Request