VIDEO ANALYSIS FOR WEAPON DETECTION AND ALERTING

Abstract

The modern world is very concerned with people's safety and security. A country's ability to attract 
investors and tourists depends on its ability to provide a problem-free environment for both groups. But 
closed-circuit television (CCTV) cameras are being used for monitoring and to record activities like 
burglary, but these cameras really need human intervention and supervision. As a result, we are looking 
for a system that can detect these illicit activities. Weapon discovery remains a significant challenge 
even with state-of-the-art deep learning calculations, fast handling equipment, and high-level CCTV 
cameras. The difficulty of the test is amplified when one considers point contrasts and the obstacles 
posed by both the gun carrier and others in the vicinity. The focus of this effort is to provide a secure 
environment where harmful weapons may be identified using CCTV footage and state-of-the-art open- 
source deep learning algorithms. With the goal of reducing misleading positives and false negatives, we 
have implemented paired order accepting gun class as the reference class and considered major disorder 
objects. Due to the lack of a universally applicable dataset, we created our own by photographing 
weapons using our own camera, manually collecting images from the web, extracting data from 
YouTube CCTV recordings, utilising GitHub stores, consulting with the College of Granada, and 
consulting the Web Films Guns Data set (IMFDB) at imfdb.org. Sliding window/order and location 
proposal/object identification are two examples of the methods used. Methods such as VGG16, 
Initiation V3, Origin ResnetV2, SSDMobileNetV1, FRIRv2, YOLOv3, and YOLOv4 are used in part. 
All of these computations were conducted with an eye on accuracy and review, as they are more 
important than precision when object identification is being done. With a mean normal accuracy that is 
91.73% greater than what was previously achieved and an F1-score of 91%, Yolov4 stands out above 
all the other computations.