Table 4. Violence detection techniques using SVM.

Ye et al. (2020)	A Video-Based DT– SVM School Violence Detecting Algorithm	Motion Co-occurrence Feature (MCF)	Optical flow extraction	Crowded	97.6%
Zhang et al. (2016)	GMOF framework with tracking and detection module	Gaussian Mixture model	OHFO for optical flow extraction	Crowded	82%–89% accuracy
Gao et al. (2016)	Violence detection using Oriented ViF	Optical Flow method	Combination of ViF and OViF descriptor	Crowded	90%
Deepak, Vignesh & Chandrakala (2020)	Autocorrelation of gradients based violence detection	Motion boundary histograms	Frame based feature extraction	Crowded	91.38% accuracy in Crowd Violence; 90.40% in Hockey dataset
Al-Nawashi, Al-Hazaimeh & Saraee (2017)	Framework includes preprocessing, detection of activity and image retrieval. It identifies the abnormal event and image from data-based images.	Optical flow and tempora difference for object detection CBIR method for retrieving images.	Gaussian function for video future analysis	Less crowded	97% accuracy
Kamoona et al. (2019)	Sparsity-Based Naive Bayes Approach for Anomaly Detection in Real Surveillance Videos	Sparsity-Based Naive Bayes	C3D feature extraction	Both crowded and uncrowded	64.7% F1 score; 52.1% precision; 85.3% recall in UCF dataset
Song, Kim & Park (2018)	SGT-based and SVM-based multi-temporal framework to detect violent events in multi-camera surveillance.	Late fusion	Multi-temporal Analysis (MtA)	Variety fight scenes from minimum two to maximum fifteen people include various movements	78.3% (SGT-based, BEHAVE), 70.2% (SVM-based, BEHAVE), 87.2% (SGT-based, NUS–HGA), and 69.9% (SGT-based, YouTube)
Vashistha, Bhatnagar & Khan (2018)	An architecture to identify violence in video surveillance system using ViF and LBP	Shape and motion analysis	ViF and Local Binary Pattern (LBP) descriptors	Both crowded and non-crowded scenes	89.1% accuracy in Hockey dataset, 88.2% accuracy in Violent-Flow dataset