Keyframe Extraction Algorithms For Bus Surveillance Video

In order to improve urban traffic condition,most countries are committed to the development of public transport system(PTS).However,the surveillance video caused by PTS makes it difficult to store and retrieve.So it becomes particularly necessary to eliminate redundant frames from these surveillance videos by keyframe extraction.Obviously,it will bring great convenience to storage and retrieval.However,it is a challenging task to achieve maximum reduction of video frames on the basis of retaining the value of data.It requires that keyframe extraction algorithm be able to accurately distinguish between keyframes and redundant frames.For this purpose,this thesis analyzes the specific characteristics and requirements of the bus monitoring system and tries a series of image processing and pattern recognition algorithms.Combined with the latest progress in the field of target detection and face recognition,a series of keyframe extraction algorithms for bus scenes are designed.The main work and results can be summarized as follows:(1)This thesis analyzes the demands of bus monitoring system and presents a complete set of keyframe extraction schemes.(2)In order to extract keyframes from the surveillance video interior of carriage,this thesis studies a variety of frame difference and background subtraction algorithms,then proposes an iterative frame difference method for keyframe extraction.(3)This thesis studies some image features such as Histogram of Oriented Gradient(HOG),Haar-like feature and classification algorithms such as Support Vector Machine(SVM),Adaptive Boosting(Adaboost).Then they are applied to head and face detection to extract the keyframes containing head or face.(4)The end-to-end target detection algorithm—“You Only Look Once”(YOLO v2)is studied in this thesis.And head and face detection models are trained based on it.(5)In this thesis,face recognition algorithms such as VGGFace,Light CNN and Sphereface are compared and analyzed from the aspects of network structure,loss function and training dataset.And based on this,an improved algorithm is proposed and achieves 99.62% on Labeled Faces in the Wild(LFW).(6)A complete set of keyframe extraction test sets are built from the existing bus monitoring data.And through the comparison experiments on these test sets,the advantages and disadvantages of above algorithms are analyzed.Finally,a set of keyframe extraction algorithms suitable for bus surveillance video are completed.Among them,the keyframe extraction algorithm based on iterative frame difference method achieves a concentration rate of about 20% in the test.And the average F-score of keyframe extraction algorithms based on head detection,face detection and face recognition reach 97.18%,99.61%,64.82% respectively in the test,fully meet the needs of practical application.