Research On Surveillance Video Anomaly Detection Algorithm Based On Multiple Probabilistic Inference In Cloud Computing Environment

The detection of abnormal events in surveillance video is a hotspot in the current research field of intelligent surveillance technology.It plays an important role in many practical applications,such as ensuring the safety of public places and maintaining traffic order.This technology can greatly save the cost of manually observing surveillance videos and reduce safety hazards caused by missed detection of manual methods.As an important research content of computer vision,the main goal of abnormal event detection is to automatically detect abnormal events in surveillance videos via computers and issue alarms.In the process of abnormal event detection,we firstly need to process surveillance video data and understand the actions and behaviors in surveillance video.Then we should make judgment for these actions and behaviors to detect the abnormal events and locate the area where the abnormal events take place.This thesis studies video abnormal event detection models and algorithms based on multiple probabilistic inference in surveillance videos.In addition,in order to accelerate the training process of abnormal event detection models for large-scale surveillance video data,this thesis realizes the parallelization of surveillance video feature extraction and multiple probabilistic inference model training algorithms via Spark cloud computing platform.The main innovations of this thesis are as follows:1.Aiming at the abnormal event detection technology in surveillance video,this thesis reviews the most important research achievements in recent years,analyzes the characteristics and applicable scenarios of various feature extraction methods in surveillance video,compares the advantages and disadvantages of different types of abnormal event detection models,summarizes 4 kinds performance evaluation criteria of the video anomaly detection methods.2.Propose a video anomaly detection algorithm based on multiple probabilistic inference(MPI-VAD).For some complex surveillance scenarios,I design a fast and effective video feature extraction method and select different model combinations according to the characteristics of each scene.I comprehensively consider the abnormal event detection results from multiple probabilistic inference models and detect different kinds of abnormal events in the surveillance video.The experiment results on 5 public datasets indicate that the proposed video abnormal event detection algorithm based on multiple probabilistic inference(MPI-VAD)has remarkable advantages in detection accuracy and real-time performance.3.Propose a parallel parameter estimation algorithm of bag-of-words model for training multiple probabilistic inference models and implement the parallel algorithm via Spark cloud computing platform.The experiment results indicate that the parallel parameter estimation algorithm of bag-of-words model greatly shortens the spent time of probabilistic inference model training;improves the efficiency of video abnormal event detection model training and has a high speedup.4.Propose a parallel parameter estimation algorithm of Finite State Markov Chain(FSMC)for training multiple probabilistic inference models and implement the parallel algorithm via Spark cloud computing platform.The experiment results indicate that the efficiency of the parallel parameter estimation algorithm of FSMC is significantly higher than that of the serial algorithm and has good scalability,which is suitable for large-scale surveillance video data.