Crowd and Mob Detection: Multi-Modal Public Safety Analytics
Multi-modal public safety analytics system combining YOLO multi-class detection, DeepSORT tracking, depth estimation, and heatmap smoothing for crowd monitoring, threat detection, and weapon-aware escalation.
Internal R&D / Personal Engineering Project at Bondstein Technologies Ltd.
Workflow at a glance
Detect people and objects
Identify visible scene entities.
Track and estimate depth
Associate tracks and relative scene depth.
Analyze proximity
Aggregate density and interaction cues.
Review risk signals
Flag events for human assessment.
Project facts & reported results
Implementation fact
Implementation fact
Implementation fact
Project Overview
Problem Statement
Public safety monitoring required real-time detection of crowd densities, people, interactions, and visible weapon-like objects across large areas. Traditional surveillance systems lacked depth awareness and intelligent escalation capabilities.
Approach & Methodology
Developed YOLO-based multi-class detection pipeline for person and object categories with interaction cues. Integrated DeepSORT tracking with Depth-Anything-V2 for monocular depth estimation. Implemented depth-aware spatial clustering, heatmap smoothing for crowd density, and weapon-aware escalation logic with person-track association across frames.
Outcome & Scope
Built an R&D pipeline for multi-class tracking, depth-aware proximity analysis, crowd heatmaps, and weapon-proximity escalation.
System Architecture
Multi-modal detection pipeline with depth-aware reasoning and escalation logic
Key Components:
Detects people and visible object categories in real-time video streams
Maintains persistent object IDs across frames for continuous tracking
Monocular depth estimation for spatial awareness and distance reasoning
Groups detections using depth information for accurate crowd density analysis
Produces smoothed density heatmaps for crowd visualization and hotspot detection
Weapon-aware threat escalation with person-track association and alert generation
Key Features & Capabilities
Multi-class detection (person/person/interaction/weapon)
DeepSORT tracking with persistent IDs
Monocular depth estimation with Depth-Anything-V2
Depth-aware spatial clustering for crowd analysis
Heatmap smoothing for density visualization
Weapon-aware escalation and person-track association
Real-time threat level assessment
Current Scope & Limitations
Detection categories and proximity cues cannot establish intent, criminality, or victim status. Alerts need human review.