Suspicious Activities Recognition in Video Sequences Using DarkNet-NasNet Optimal Deep Features  

作　　者：Safdar Khan Muhammad Attique Khan Jamal Hussain Shah Faheem Shehzad Taerang Kim Jae-Hyuk Cha 

机构地区：[1]Department of Computer Science,COMSATS University Islamabad,Wah Campus,Wah Cantt,47040,Pakistan [2]Department of Computer Science,HITEC University,Taxila,47080,Pakistan [3]Department of Computer Science,Hanyang University,Seoul,04763,Korea

出　　处：《Computer Systems Science & Engineering》2023年第11期2337-2360,共24页计算机系统科学与工程（英文）

基　　金：supported by the“Human Resources Program in Energy Technology”of the Korea Institute of Energy Technology Evaluation and Planning (KETEP)and granted financial resources from the Ministry of Trade,Industry,and Energy,Republic of Korea (No.20204010600090).

摘　　要：Human Suspicious Activity Recognition(HSAR)is a critical and active research area in computer vision that relies on artificial intelligence reasoning.Significant advances have been made in this field recently due to important applications such as video surveillance.In video surveillance,humans are monitored through video cameras when doing suspicious activities such as kidnapping,fighting,snatching,and a few more.Although numerous techniques have been introduced in the literature for routine human actions(HAR),very few studies are available for HSAR.This study proposes a deep convolutional neural network(CNN)and optimal featuresbased framework for HSAR in video frames.The framework consists of various stages,including preprocessing video frames,fine-tuning deep models(Darknet 19 and Nasnet mobile)using transfer learning,serial-based feature fusion,feature selection via equilibrium feature optimizer,and neural network classifiers for classification.Fine-tuning two models using some hit and trial methods is the first challenge of this work that was later employed for feature extraction.Next,features are fused in a serial approach,and then an improved optimization method is proposed to select the best features.The proposed technique was evaluated on two action datasets,Hybrid-KTH01 and HybridKTH02,and achieved an accuracy of 99.8%and 99.7%,respectively.The proposed method exhibited higher precision compared to existing state-ofthe-art approaches.

关 键 词：Suspicious activity deep learning WEIGHTS SURVEILLANCE optimization 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]