| With the rapid development of computer hardware and artificial intelligence technology,human behavior recognition has been favored by a large number of researchers due to its high theoretical research value and wide application in daily life.It has not only been constantly innovated at the theoretical level,but has also been widely applied in intelligent monitoring,human-machine interaction,medical rehabilitation,safety and elderly care,and sports.Early human behavior recognition based on handcrafted features often required artificial design of features and judgment rules,which not only consumed time and energy,but also had a bottleneck in recognition accuracy.In recent years,deep learning technology has been widely used in the field of human behavior recognition.The deep neural network avoids the tedious process of handcrafted feature design and quantization and replaces it with more excellent deep features,achieving better recognition effect.However,there is still a lot of room for improvement in deep learning methods in terms of sample scale normalization,small sample learning,multi-scale learning,and further processing of extracted features.Taking computer vision-based human behavior recognition as an example: first,the duration of each behavior video is different,and cannot be trained and tested as the same batch of data in the model;second,the number of samples in a single dataset is not enough to meet the training needs of the model,making it difficult to truly demonstrate the role of deep learning models,and easily leading to overfitting;finally,even if the model extracts a large number of features,there are differences between the features,and not all features have the same effect on the current task,so it is necessary to distinguish the weight coefficients of the added features to ensure that the model’s limited attention and the computer’s limited computing power are concentrated on the part of the features with greater weight.In response to the above problems,this article used the classical convolutional neural network model as the framework and carried out the following research work.:To address the problem of insufficient training samples leading to overfitting in the network model.This paper proposes a data augmentation strategy based on human skeleton.Although each person’s appearance and body shape are different during the execution of a behavior,the underlying human skeleton remains roughly the same.Therefore,this paper scales the skeleton of the behavior executor within a certain range to generate new behavior samples and quantify them using a three-dimensional matrix representation.The experimental analysis on the Florence-3D and UT-3D datasets shows that this strategy effectively solves the problem of model overfitting and greatly improves the accuracy of human behavior recognition.To address the problem of inconsistent time dimensions and inability to batch train human behavior samples in the dataset,this paper proposes an efficient scale normalization strategy that preserves motion information.Firstly,the duration of all samples of a single behavior is statistically analyzed,and the normalization scale is set at the median or nearby.Secondly,through linear interpolation fitting and frame skipping strategy,all samples complete the scale normalization operation while maximizing the preservation of the original information of the action.To address the issue that deep learning models struggle to further distinguishing and processing the rich motion information contained in action samples after feature extraction.First,the combination of spatial pyramid pooling layer and global average pooling layer based on the spatial pyramid pooling layer gives the model the ability to perform feature learning on different scale samples.Then,attention modules are embedded to adaptively adjust the importance of channels and spatial regions,thus improving the feature representation ability of the model.This allows the deep model to selectively focus limited attention on the features that have a greater impact on the task of human behavior recognition,while suppressing excessive attention on nonimportant features,thereby improving the accuracy of the model in recognizing different human postures and actions. |
