基于联合轨迹特征的徘徊行为识别方法

Wandering recognition method based on joint trajectory features

徘徊行为判定作为异常行为识别的热点难题,目前存在难以构建具有场景特定性的识别模型问题.特定场景下的行人轨迹数据,或是有标签,或是无标签,针对如何利用特定场景下有标签轨迹数据信息问题,结合行人多种轨迹特征,提出基于分类函数的徘徊行为识别模型,能够通过分类函数自动学习该场景下有标签轨迹数据中的徘徊行为模式;针对如何利用特定场景下无标签轨迹数据信息问题,提出基于异常检测的徘徊行为识别模型,能够在大量无标签轨迹数据中自动学习潜在的徘徊行为模式.基于两种徘徊行为识别模型,提出徘徊行为识别框架,能够利用目标跟踪算法获取特定场景视频中行人的轨迹数据,并根据数据是否带有标签合适地构建对应的具有场景特定的徘徊行为识别模型.为了验证所提模型的有效性,选用CASIA⁃AR公开数据集作为测试集,和三种基准模型一起,在行走、奔跑和徘徊行为的识别上进行了对比实验.实验结果表明,所提模型在测试集上的准确率和召回率都优于基准模型,F1指标也有提升,验证了所提模型的有效性、场景特定性和迁移性.

Wandering behavior determination is hot in abnormal behaviors recognition problem.At present,it is difficult to construct a scene⁃specific recognition model.Pedestrians'trajectory data from specific scene may be labeled or unlabeled.In view of the problem of how to use the information of labeled trajectory data from a specific scene,this paper proposes a novel wandering behavior recognition model based on classification functions by combining multiple trajectory features of pedestrians,which can automatically learn the wandering behavioral patterns from labeled trajectory data from the specific scene through the classification functions.In view of the problem of how to use the information of unlabeled trajectory data from a specific scene,this paper proposes a novel wandering behavior recognition model based on outlier detection,which can automatically learn potential wandering behavioral patterns from a large amount of unlabeled trajectory data from the specific scene.Based on these two novel wandering behavior recognition models,this paper proposes a wandering behavior recognition framework which uses object tracking algorithm to obtain pedestrians'trajectory data from videos of specific scene,and constructs scene⁃specific wandering behavior recognition model according to whether the data has labels or not.We use CASIA⁃AR public dataset as the test set,and verify the effectiveness,scene specificity and transfer ability of the proposed model by comparing the recognition result on walking,running and wandering behavior with three benchmark models.Experimental results show that the proposed models have high precision and recall on the test set,yielding an improvement F1 score compared with three benchmark models.