Study And Design Of Understanding Actions And Detecting Anomaly Actions For Typical Targets

The paper takes tanks and airplanes as typical targets,and studies the method of understanding actions made by typical targets and detecting anomaly actions both in videos.This study aims to infer actions made by typical targets through analyzing acting rules of targets in obtained videos,and classify anomaly actions based on moving trajectories of typical targets.Issue occurs in the current video motion target detection methods is the imcomplete area detection of moving target in video due to hole phenomenon,which may cause high false detection rate.In order to solve this issue,this paper proposes a video target detection method based on the improved motion template and histogram matching.The method uses three-frame differences and the adaptive thresholding method to detect the contour of instead of using the target contour detection method in the traditional motion template method.In addition,this paper uses the temporal segmentation method to segment the video frame images to detect the potential region of moving targets in videos.Then,this paper adopts the HSV histogram matching method to clarify the detected potential target areas,by eliminating falsely detected areas to improve the detection accuracy.In addition,this paper proposes an action inference method based on the motion states of targets in videos.This method first detects a typical target,together with its launch states,motion speed states and environment states.Then this paper adopts optimization-based junction tree algorithm to infer the action type of targets,to improve prediction accuracy.The experimental results have shown that the proposed method can obtain better results in detecting of typical targets than the background subtraction,guided filtering and mixed gaussian models and motion template method and infering the action types according to the motion states of the typical targets.Issue occurs in the current trajectory prediction methods based on deep learning techonologies has low prediction accuracies without considering the interaction relationships between trajectories of moving targets.This paper proposes a trajectory prediction method based on social long-term memory network.In predicting trajectories of moving targets,the influences of moving trajectories of adjacent targets at the same frame are taken into account by increasing the S-Pooling layer.Then the paper calculates the abnormality according to the predicted trajectories of typical targets across videos,together with weather,location and anomaly indicator libraries,to classify abnormal types and grades.And a alert sound is giving when a classified abnormal type with a high risk grade.In the detection of the trajectories of targets,this paper proposes an improved TLD algorithm,which combines the prediction function of the Kalman algorithm on the basis of the TLD algorithm,so that a video with multiple targets and trajectories can be classified correctly.It has the potential to solve the problem when using TLD algorithm which may miss detecting trajectories due to losing of targets.The experimental results have shown that the proposed method can obtain targets motion trajectories better,and obtain trajectory integrity higher than LK and KCF.