Research On Parallel Algorithm Of Detection And Tracking Mutiple Vehicles And Persons From Surveilliance Video Based On Hadoop Clusters

Video surveillance in today's society has been widely used.With the implementation of the national "Safe City" policy,more video surveillance equipment will be installed in the future.An important purpose of installing video surveillance equipment is to ensure social safety and effectively prevent and detect crimes and criminal activities.After the illegal criminal incident,the surveilliance video will become an extremely important tool for police investigators to solve the case.However,up to date,public security criminal investigators use almost entirely their naked eyes to watch these surveillance videos.It can be said that the efficiency is very low.Therefore,this thesis studies the method of parallel detection and tracking of peoplevehicle target in surveillance video based on Hadoop cluster.Following research work have been done in this thesis:(1)Since frame difference method has high parallelism,it is suitable to implemented on Hadoop clusters.As for holes are often produced by frame differencing,this thesis proposed an effective hole-filling method to solve the problem.The frame difference method is thus improved and get better results.(2)According to the parallel mechanism of Hadoop,a splitting method has been designed for splitting surveillance video into video segments so that each video segment is put into a Hadoop map task to segment moving objects using the improved frame difference method proposed in this paper.So segmentation of moving objects in the video is implemented at the video segment level.Thus the efficience of segmentation of entire video is greatly improved on Hadoop clusters by compared with stand-alone PC.The segmentation of moving vehicles and persons is the basis for tracking vehicles and persons.Since tracking an object is time ordering,proper keys are designed to transform from Map to Reduce to ensure that all the results of the maps to transfer to Hadoop Reduce correctly in time order.(3)In order to track the long-range Vehicles and Persons moving target in the video and overcome the lack of local features such as clear face and license plate number,this paper adopts the color histogram as the feature of the moving target,Histogram features of 8 × 8 × 8 levels were extracted in color space.(4)Since the color histogram has no obvious characteristic significance for very small area,this paper studies the generation of particles in the particle tracking algorithm,improves the generation of particles,greatly reduces the number of particles,but also guarantees the particle Which not only avoids the particle tracking algorithm from falling into local convergence but also improves the computational efficiency of the algorithm.(5)There is a time-dependent dependency on the tracking of a moving object in the whole video.This paper studies the rationality of tracking moving objects in Reduce Hadoop,designs and implements a method based on the segmentation of moving objects,The multi-objective and improved particle filter tracking algorithm avoids tracking the moving vehicle to a similar vehicle parked on the same vehicle,and improves the computational efficiency of the particle tracking algorithm through improved particle generation and evolution.(6)The entire Hadoop cluster-based surveillance video multi-motion target tracking experimental prototype system was implemented.The algorithm developed in this paper can provide the support for improving the efficiency of video crime investigation by public security criminal investigators.