| Duo to the information complexity of the tranditional color image and the influence of the illumination change,smog and the poor range of visibility in dark environments,the abnormal behavior detection based on the color videos has poor results and low frame rate.However,the thermal infrared imager can work effectly when facing the problem of lack of nighttime visibility,sudden illumination variation and hazy weather.Therefore,the videos is collected by the FLIR FC-S series thermal infrared imager and based on the collected thermal videos,the abnormal behaviors are detected.The main contents of this paper are as follows:1.This thesis employs a conrour-based background-subtraction algorithm to extract foreground objects from thermal videos.The main steps are: using statistical background-subtraction technique to identify local regions-of-interest,building and thinning the contour saliency map,using watershed transform method to complete contour line,lastly,adopt Flood-Fill method to produce foreground object silhouettes.2.This thesis presents a cell-based abnormal behavior detection algorithm.As the abnormal behavior has low occurrence probability,short duration and aperiodicity,it is difficult to model abnormal behavior.In this paper,we use the following strategy:Firstly,we extract the foreground objects from training videos and testing videos.Secondly,we divide the foreground objects into cells and extract the features from each cell.Thirdly,we build normal behavior models based on features of trainning videos.Lastly,we compare the features of cells of testing videos to normal behavior model.If comparison result is similar,the behavior in this frame is judged as normal behavior.Otherwise,judged as abnormal behavior.Finally,we do experiments on 14 videos including different abnormal behaviors.The experiment results show that our method can recognize and track the abnormal behaviors and the frame rate can stay around 11 fps that can satisfy the needs of practical applications. |
PDF Full Text Request