| Currently,video has emerged as a primary information carrier on the Internet,following text and images.The generation of massive amounts of data has led to the widespread application of video data analysis tasks.Among various video understanding tasks,action recognition has become a core component of video data analysis businesses due to its ability to integrate temporal and spatial information.The accuracy and real-time performance of action recognition algorithms directly impact the operation of upper-layer business.Although current action recognition algorithms have been successfully applied in various fields,their application in edgeside deployment scenarios still presents numerous challenges.On one hand,in order to meet the increasing demand for model accuracy,models are becoming increasingly complex,resulting in a significant increase in model parameters and computational operations.On the other hand,the limited hardware resources pose hurdles to deploying such intricate models,characterized by a high number of parameters and extensive computational requirements,on edge-side devices.This study investigates and explores the action recognition edge algorithm and deployment system based on deep neural networks to address the aforementioned challenges.The following achievements are obtained.Firstly,considering the issues in modeling temporal domain information using current 3D convolutional networks,this paper proposes a lightweight temporal global attention module.Experimental results demonstrate that by inserting this module into the target network,the modified model can effectively improve accuracy with minimal increase in parameters and floating-point operations,partially mitigating the issue of escalating model complexity that arises from improving model accuracy.Secondly,this paper studies the model pruning algorithm applied to 3D convolutional networks for action recognition and proposes a pruning algorithm based on the average activation value of the Region of Interest(RoI)region in the feature map.This algorithm effectively eliminates parameter redundancy in the action recognition model by using the average activation value of the feature map in the region where the character is located as an evaluation indicator to measure the importance of the filter.Additionally,a hybrid pruning strategy for the SlowFast network,known for its outstanding performance in 3D convolutional networks for action recognition,is introduced.By combining the proposed pruning algorithm with this hybrid strategy,the pruned model achieves higher accuracy than other commonly used pruning algorithms.Finally,this paper designs and implements a parallel inference framework and deployment system for edge-side devices,helping developers easily deploying algorithmic models to edge-side devices for efficient inference by leveraging a modular structure and a purposeful design. |
