VIDEO ANALYSIS FOR WEAPON DETECTION AND ALERTING

Abstract

The welfare of individuals is of paramount importance in today's society. If a nation can make its 
visitors and investors feel safe, they will be able to draw in more of both types of visitors and investors. 
The usage of closed-circuit television (CCTV) cameras to capture and monitor events like burglaries 
highlights the fact that these cameras rely heavily on human oversight and intervention. Therefore, we 
are in need of a system that can identify such illegal actions. The use of cutting-edge deep learning 
algorithms, lightning-fast handling gear, and top-tier CCTV cameras has not solved the problem of 
weapon finding. When using point contrasts and other nearby impediments, such as the gun carriage, 
the test becomes much more challenging. With the use of cutting-edge open-source deep learning 
algorithms and CCTV data, this endeavour aims to create a safe space where dangerous weapons may 
be detected. Using the gun class as a reference and significant disorder objects as our focus, we have 
introduced paired order acceptance in an effort to decrease false negatives and misleading positives. We 
built our own dataset by taking pictures of weapons with our own camera, manually searching the web, 
using GitHub stores, consulting with the College of Granada, and extracting data from YouTube CCTV 
recordings. We also consulted the Web Films Guns Data set (IMFDB) at imfdb.org. Two examples of 
the strategies utilised are sliding windows/order and location proposals/object identification. Some of 
the methods that are employed are VGG16, Initiation V3, Origin ResnetV2, SSDMobileNetV1, 
FRIRv2, YOLOv3, and YOLOv4. Precision is less critical than accuracy when it comes to item 
recognition, therefore that's why we ran all these calculations with review in mind. Yolov4 is the most 
impressive calculation since its mean normal accuracy is 91.73% higher than the previous result and it 
has an F1-score of 91%.