Research On Accurate Extraction And Evaluation Mechanism Of Surveillance Video Key Frames

With the improvement of the public awareness of security,millions of surveillance cameras are installed,resulting in explosive growth of surveillance video data.It has become a tricky issue for people to implement efficient management,fast retrieval and query of surveillance video data.As an effective method to address this issue,key frame extraction has attracted the attention of researchers.So far,key frame extraction has achieved many research results.However,the results are not accurate when existing effective key frame extraction methods are directly applied to surveillance video.Compared with the rapid development of video key frame extraction methods,video key frame evaluation methods have developed relatively slowly.Few objective evaluation methods are used to appraise video key frames.And the widely used fidelity,shot reconstruction degree,precision and recall,etc.,are not suitable for evaluating key frames of surveillance video.Therefore,this dissertation mainly studies the accurate extraction and evaluation mechanism of surveillance video key frames.Major research contents are as follows:1.Accurate key frame extraction of surveillance videoPeople pay more attention to the changes of objects motion state in surveillance video.Therefore,the motion state changes of objects in video can more accurately reflect the change of the monitored video content.Therefore,extracting video frames,which include the motion state changes of objects,can reflect the changes of surveillance video content.In order to accurately capture the motion state changes of objects and extract the key frames of surveillance video,the following three methods are proposed:(1)A key frame extraction method of surveillance video based on motion velocity is proposed.Compared with existing key frame extraction methods,this method uses motion velocity to describe the motion state changes of objects in video.Firstly,the mainly moving object in video is extracted by moving objects detection,and the coordinate position of object is obtained.Secondly,the moving velocity of object in each frame is calculated.Considering the importance of changing direction of motion,moving and still,the absolute of velocity is weighted.Then the velocity curve is formed according to the processed velocity,and its peak point is detected.Finally,the video frames at the peak mutation are extracted as key frames.Experimental results have shown that the proposed method compared with the contrast methods can accurately and timely capture the changes of object's motion state.And experimental results verify the correction and effectiveness of the proposed method.(2)A key frame extraction method of surveillance video based on center offset is proposed.As the key frame extraction method of surveillance video based on motion velocity will miss the changes of local motion state of objects,the center offset is proposed to capture the changes of moving objects global and local motion state.Firstly,the moving objects are extracted by moving objects detection.The center point of the object moving shape is used to replace the object.When there are multiple objects in video frame,the mean value of the coordinates of multiple center points is calculated as the center point of the video frame.Next,the center offset of each frame is calculated.Then the center offset of each frame is connected to form the center offset curve.Finally,Video frames at the peak mutation are extracted.And the final key frames are determined by employing a visually distinguishing mechanism to further simplify the extracted video frames.The experimental results demonstrate that the proposed method outperforms the existing state-of-the-art method in terms of shot reconstruction degree and capturing the local motion state changes of moving objects.(3)A key frame extraction method of surveillance video based on frequency domain analysis is proposed.Compared with key frame extraction method based on center offset,this method analyzes video frames in the frequency domain,and utilizes the changes of frequency domain information to reflect the changes of objects motion state.Firstly,the two-dimensional Fourier transform is applied to video frames to obtain the spectrum and phase of each video frame.Next,the mean square errors of the spectrum and phase of adjacent frames are calculated respectively.The mean square errors of the spectrum and phase are weighted.Then,the processed mean square error value is added to get the mean square error of each frame,forming the mean square error curve.Finally,Video frames at the peak mutation are extracted.And the final key frames are determined by utilizing a visually distinguishing mechanism to further simplify the extracted video frames.Experimental results have shown that the proposed method is superior to the contrast methods in shot reconstruction degree and can more accurately capture the changes of local motion state of objects.2.Key frame evaluation mechanism of surveillance videoIn surveillance video,video frames whose motion state changes of objects are more important,so the method,which is suitable for evaluating the key frames of surveillance video,needs to evaluate the extracted key frames from capturing motion state changes of the object.Aiming at the evaluation needs of surveillance video key frames,starting from the object motion trajectory,an evaluation mechanism suitable for the key frames of surveillance video,called motion trajectory reconstruction,is proposed.Firstly,the evaluation mechanism employs moving object detection and tracking to extract the moving trajectory of each object in video.And the original moving trajectory set of each object is formed.Then,according to each object motion trajectory of the extracted key frames,a linear interpolation algorithm is used to reconstruct each object motion trajectory set.Finally,the original motion trajectory set is compared with the reconstructed motion trajectory set,separately.The higher the degree of coincidence,the stronger the ability of the key frame extraction method to capture the changes of object's motion state.The experimental results show that the proposed evaluation mechanism can evaluate the key frames of surveillance video from capturing the global and local motion state changes of objects.