Assessment of Data Augmentation Techniques for Firearm Detection in Surveillance Videos

Rose, J. Dafni; Bourlai, Thirimachos; Loudermilk, James A.

doi:10.1109/bigdata50022.2020.9378267

Cited by 5 publications

(2 citation statements)

References 17 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…In the unmatched Gaussian distributions, the 𝜇 and 𝛴 parametric values are similar, and the matched parameters of Gaussian 𝐺 𝑖 in the mixture 𝑋 𝑡 is updated as mentioned in the Eqs. ( 6) to (8).…”

Section: Parameter Estimation Of K-gaussian Distributionmentioning

confidence: 99%

“…Usually, the human operator handles the weapon detection task, which is ineffective, due to visual distraction or fatigue [6,7]. In addition, the increasing number of areas controlled by video cameras and the factors inherent to human conditions like loss of attention and fatigue make these systems inefficient [8]. Therefore, intelligent systems are developed by researchers for the automatic detection of risk situations or threats involving firearms [9,10].…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Weapon Detection in Surveillance Videos Using Human Inspired Particle Swarm Optimization Algorithm and Support Vector Machine

2022

IJIES

View full text Add to dashboard Cite

In recent decades, automatic control systems are becoming the essential need for security forces, due to the increase in the number of criminal activities. The fast and precise automatic weapon detection system is useful to mitigate or avoid risks in public spaces. In this manuscript, a new automated model is implemented for effective weapon detection in closed circuit television videos. After collecting the data from YouTube and Gun movies databases, the Gaussian Mixture Model (GMM) is applied to detect the weapons in the video sequences. Then, the feature extraction is performed using deep learning models: AlexNet and ResNet 18, and a descriptor: Scale Invariant Feature Transform (SIFT) for extracting the feature vectors from the segmented regions. Whereas, the combination of deep and texture features reduces the semantic space between the feature sub-sets that helps in enhancing the classification performance. In addition, the feature optimization is accomplished by Human Inspired Particle Swarm Optimization (HIPSO) algorithm to select active feature vectors that decrease the system complexity and training time of the classifier. In the conventional PSO algorithm, the Human Group Optimization (HGO) algorithm is utilized to influence the particles, and then the adaptive uniform mutation is utilized to improve the convergence rate and makes the implementation simple. Finally, the selected active feature vectors are fed to the Support Vector Machine (SVM) classifier for weapon and non-weapon classification. The experiment results confirmed that the HIPSO-SVM model has achieved high accuracy of 95.34% and 98.60% on the YouTube and Gun movies databases, which are better compared to the existing models.

show abstract

Section: Parameter Estimation Of K-gaussian Distributionmentioning

confidence: 99%