[Paper Review] Knife and Threat Detectors

David Noever, Sam E. Miller Noever|arXiv (Cornell University)|Apr 4, 2020

Forensic and Genetic Research25 references4 citations

TL;DR

This paper proposes a multi-method framework for automated knife threat detection using computer vision and deep learning, leveraging MobileNet for efficient classification (95% accuracy, <2.2 MB memory), Mask R-CNN for hand-knife localization, and PoseNet for anatomical threat inference. The key contribution is a systematic, low-resource pipeline for real-time threat detection in surveillance, minimizing false positives and negatives with curated negative training data.

ABSTRACT

Despite rapid advances in image-based machine learning, the threat identification of a knife wielding attacker has not garnered substantial academic attention. This relative research gap appears less understandable given the high knife assault rate (&gt;100,000 annually) and the increasing availability of public video surveillance to analyze and forensically document. We present three complementary methods for scoring automated threat identification using multiple knife image datasets, each with the goal of narrowing down possible assault intentions while minimizing misidentifying false positives and risky false negatives. To alert an observer to the knife-wielding threat, we test and deploy classification built around MobileNet in a sparse and pruned neural network with a small memory requirement (&lt; 2.2 megabytes) and 95% test accuracy. We secondly train a detection algorithm (MaskRCNN) to segment the hand from the knife in a single image and assign probable certainty to their relative location. This segmentation accomplishes both localization with bounding boxes but also relative positions to infer overhand threats. A final model built on the PoseNet architecture assigns anatomical waypoints or skeletal features to narrow the threat characteristics and reduce misunderstood intentions. We further identify and supplement existing data gaps that might blind a deployed knife threat detector such as collecting innocuous hand and fist images as important negative training sets. When automated on commodity hardware and software solutions one original research contribution is this systematic survey of timely and readily available image-based alerts to task and prioritize crime prevention countermeasures prior to a tragic outcome.

Motivation & Objective

• To address the under-researched problem of automated knife threat detection in video surveillance despite high annual assault rates.
• To reduce false positives and false negatives in identifying potentially dangerous knife-wielding individuals using image-based machine learning.
• To develop a lightweight, deployable system suitable for commodity hardware with minimal memory footprint.
• To identify and address data gaps by collecting innocuous hand and fist images as negative training examples.

Proposed method

• A MobileNet-based sparse and pruned neural network is trained for fast, accurate classification of knife threats with under 2.2 MB memory usage.
• A Mask R-CNN model is employed to segment and localize both the hand and knife in a single image, enabling relative position analysis for threat assessment.
• PoseNet is used to detect anatomical waypoints, inferring threat posture and intent through skeletal feature analysis.
• The system integrates multiple models to improve threat scoring by combining classification, localization, and pose-based reasoning.
• Negative training data, including images of hands and fists in non-threatening poses, are collected to reduce false positives.
• The framework is designed for deployment on commodity hardware using off-the-shelf software, ensuring practical real-world applicability.

Experimental results

Research questions

• RQ1 How can a lightweight deep learning model achieve high accuracy in knife threat detection while maintaining low memory usage?
• RQ2 What role does hand-knife spatial localization play in improving threat assessment accuracy?
• RQ3 Can pose estimation from skeletal features enhance the interpretation of threatening intent?
• RQ4 How do curated negative examples (e.g., hands, fists) impact false positive rates in threat detection?
• RQ5 To what extent can a multi-model system improve overall threat detection performance compared to single-method approaches?

Key findings

• The MobileNet-based classifier achieved 95% test accuracy with a model size under 2.2 megabytes, enabling efficient deployment on low-resource systems.
• The Mask R-CNN model successfully localized both knife and hand in images, enabling relative position analysis to infer overhand threat postures.
• PoseNet-based skeletal feature detection contributed to distinguishing threatening from non-threatening stances, reducing misinterpretation of intent.
• The inclusion of negative training data—such as images of hands and fists—significantly reduced false positive detections.
• The multi-model framework demonstrated improved threat detection performance by combining classification, localization, and pose analysis.
• The system is deployable on commodity hardware using standard software stacks, enabling practical real-time surveillance integration.