International Research Journal of Engineering and Technology (IRJET) e-ISSN: 2395-0056 Volume: 13 Issue: 02 | Feb 2026
www.irjet.net
p-ISSN: 2395-0072
Daroda Pratibandhak-An Anti Robbery System to protect shops and Banks from Robbery using Deep Learning and IoT R. S. Kakade1, Adak Rupesh Hanumant2, Labade Vivek Anil3, Shejul Devendra Shirish4 1HOD, Dept. of Computer Technology, P.Dr.V.V.P. Institute of Technology and Engineering (Polytechnic), Loni,
Maharashtra, India
2,3,4 Final year Diploma Student, P.Dr.V.V.P. Institute of Technology and Engineering (Polytechnic), Loni,
Maharashtra, India ---------------------------------------------------------------------***---------------------------------------------------------------------
Abstract - There is a direct correlation between the
width, and input resolution with its revolutionary compound scaling technique. It was chosen for this reason. This balanced scaling improves feature extraction with fewer parameters than conventional models, which solely increase layer count to improve performance. Retail and banking applications require low-latency HD video processing on commodity hardware or edge devices. To simulate a welltrained computer brain, the model meticulously evaluates each frame of the video for intricate visual patterns that indicate danger, such as the precise outlines of a weapon or the feel of a mask, while ignoring background features.
increase in youth unemployment and the exponential growth in the number of young people without the necessary skills. A rise in criminal activity, including robbery and burglary, is generally accompanied by these variables. Retailers, including jewelers and others, bear a disproportionate share of the crime rate associated with armed robberies. Offenders frequently hide their identities after an incident because they cover their faces. auther can successfully employ AI to prevent these outcomes and promptly notify the authorities in order to curb this threat. For improved accuracy, auther train the weapon dataset—which includes knives and revolvers—using a deep learning model called EfficientDet D4. The system starts receiving live feeds from the CCTV cameras after the data has been trained. The proposed model promptly activates the IOT model and issues a loud audio alert whenever a shoplifter brandishes a weapon against the store owner. When the IoT model is turned on, an Arduino-driven motor is instructed to automatically close the store's door. A loud audio alert is then raised, and a WhatsApp alert is sent to the closest police station, along with a photo of the store and its location on a map.
The project carefully blends several cutting-edge approaches to ensure the identification process works in varied situations. After pre-training on ImageNet, a successful net 4 model is fine-tuned using hundreds of weapons, masks, and suspicious stance photographs. This method focuses on transfer learning. This allows the system to focus on a problem while using its enormous library of visual attributes. Using backdrop matting and anomaly detection is another key method. Equipment can detect nonthreatening "abandoned objects" from ordinary consumer flow. Real-time object identification and multi-object tracking monitor visitors' movements throughout the facilities for hostility or excessive time in prohibited areas. The system provides comprehensive and scalable security by combining optical identification with automated alert protocols like the simple mail transmission protocol for timely notifications. It transforms static surveillance into smart, dynamic defense.
Key Words: AI, EfficientDet D4, closed-circuit television, internet of things (IoT) paradigm
1. INTRODUCTION Traditional surveillance isn't always adequate to protect financial institutions and retail locations from highstakes crimes like armed robberies in today's technologically advanced world. Traditional CCTV systems are better for forensic investigation than for protecting people or property. To overcome this issue, financial and retail institutions need an automated anti-robbery system. This digital sentry might employ AI's lightning-fast decision-making to anticipate and eliminate threats. This technology's debut is crucial since crimes can happen faster than guards or staff can activate alarms. These technologies inform authorities quickly if they detect suspicious activities or risks. This creates a constant line of defense without human fatigue or distraction, dramatically reducing response times, and ending bloodshed. The research relies on Google AI's net 4 deep learning model's robust convolutional neural network design. Its unique blend of processing speed and precision makes it famous. The efficient b4 model of net 4 scales depth,
© 2026, IRJET
|
Impact Factor value: 8.315
[1] Adding class and location labels to each training image is proposed by Keong-hun choi et al. to improve supervised learning for object detection. The ability of the new environment to identify things outside of the training context depends on the presence of a matching label. Our research presents an object detection system that utilizes reinforcement learning. A few photos and an inventory of the contents will do the trick. Three models are proposed: one for area-based evaluation, one for incentive configuration, and one for transformer-based item proposal. [2] Khalid Elgazzar and colleagues investigated the application of deep learning to object detection. Using the training capabilities of Convolutional Neural Networks (CNNs), thousands of objects can be reliably detected even in tough lighting and occlusion settings. The availability of a large number of training images
|
ISO 9001:2008 Certified Journal
|
Page 822