Real-Time Weapon Detection with TensorFlow Object Detection API

Problem Statement

Over the past decade, the frequency of mass shootings in schools has increased dramatically. Gun violence on school grounds has left many children injured or killed, reflecting a wider issue of firearm-related violence in the United States. As long-term solutions are debated, there is a growing need for immediate alternatives that can help reduce or prevent such incidents.

On average, police response time is around 18 minutes. During an active shooting, calling 911 may not be possible right away—people often need to find safety first, which delays the process. This means the 18 minutes only accounts for response time after the call, not the time it takes to make it.

Project Overview

This project explores the use of real-time weapon detection integrated with CCTV cameras to drastically reduce police response time. The object detection model is designed to identify weapons in images, videos, or live video streams.

Once a weapon is detected on a live feed, authorities can be alerted immediately. They can monitor the feed, confirm the threat, and dispatch units without relying on a phone call. This can cut down response times significantly and potentially prevent casualties.

Detection Methodology

Training an object detection model from scratch is both time- and resource-intensive. Instead, transfer learning provides a more practical approach. TensorFlow’s Object Detection API allows customization of pre-trained models, originally trained on general datasets like COCO, to detect domain-specific objects.

Some of the available pre-trained models include:

Model name	Speed (ms)	COCO mAP
ssd_mobilenet_v1_coco	30	21
ssd_mobilenet_v2_coco	31	22
ssdlite_mobilenet_v2_coco	27	22
ssd_inception_v2_coco	42	24
faster_rcnn_inception_v2_coco	58	28

For this project, I chose ssd_mobilenet_v2_coco because:

1. It provides low inference time (31 ms), suitable for real-time webcam input.
2. It offers a reasonable accuracy score (mAP).

Dataset

The dataset was sourced from the University of Granada research group. It consists of ~3,000 images of firearms (mainly handguns) in .jpg format, each with an accompanying annotation file in .xml format containing class labels and bounding box coordinates.

Example XML structure:

<annotation>
  <filename>armas (3)</filename>
  <size>
    <width>1300</width>
    <height>866</height>
    <depth>3</depth>
  </size>
  <object>
    <name>pistol</name>
    <bndbox>
      <xmin>471</xmin>
      <ymin>207</ymin>
      <xmax>613</xmax>
      <ymax>359</ymax>
    </bndbox>
  </object>
</annotation>

Training Process

The model was trained on Google Colab using GPU acceleration. The training lasted around 14 hours and involved fine-tuning hyperparameters, such as:

• Image augmentation (rotation, flipping, resizing, color adjustments)
• Learning rate modifications
• Regularization techniques (e.g., dropout)

TensorBoard was used to monitor performance metrics like loss and Mean Average Precision (mAP) in real time.

Model Evaluation

The evaluation primarily focused on mAP scores and Intersection over Union (IoU) thresholds:

A detection is considered True Positive if its IoU exceeds the chosen threshold, otherwise it’s classified as False Positive.

Conclusion

This project demonstrates how AI-powered surveillance can help reduce police response time during firearm-related incidents. While further improvements are needed to minimize false positives, the method shows strong potential as a preventive tool.

Beyond schools, such technology could be deployed in other public and private spaces—parks, banks, gas stations, or shopping centers—providing an additional layer of security.