AI-Powered Multi-Camera Person Re-Identification
We partnered with a leading security technology provider in Saudi Arabia to build a next-generation surveillance platform. By fusing Face Embeddings, Gait Analysis, and Shoe Pattern Recognition, we engineered a system capable of tracking individuals across non-overlapping camera networks with high precision, even in crowded or low-light environments.
The Impact Dashboard (Metrics)
The "Broken Track" Problem
Tracking a subject across a large facility with hundreds of cameras is notoriously difficult. Traditional systems lose the “Identity Lock” when a person walks into a blind spot or changes direction (non-overlapping fields of view). Reliance on face recognition alone failed when subjects turned away from the camera or wore masks. The client needed a robust solution that could maintain identity consistency using multiple biometric signatures.
Key Bottlenecks
Identity Drift
Subjects were often re-assigned new IDs when moving between cameras, creating fragmented tracking logs.
Occlusion Failure
Standard models failed when a person was partially blocked by a crowd or object.
Low-Light Variance
Noisy footage from older cameras degraded facial recognition accuracy significantly.
Search Latency
Matching a query image against millions of historical vectors slowed down real-time alerts.
Client Profile
Industry
Region
Saudi Arabia
Focus
Smart Surveillance
Core Tech
YOLOv10, SwinGait, InsightFace, FAISS, Docker
Multi-Modal Biometric Fusion
Inexture.ai engineered a Multi-Modal Re-ID System. Instead of relying on a single feature, we treat identity as a composite vector. We use InsightFace for facial features, SwinGait to analyze walking patterns, and a custom ResNet-50 model to extract shoe and clothing embeddings. These vectors are fused and indexed in FAISS for millisecond-level retrieval.
Engineering The Platform
Advanced Person Detection
Deployed YOLOv10 for high-accuracy bounding box detection and YOLOv10-Pose for skeletal keypoint estimation. This ensures the system captures the subject even in crowded scenes.
+22% improvement in initial detection accuracy.
Gait Recognition Pipeline
Integrated SwinGait (Swin-T), a deep learning model that identifies individuals by their unique walking cadence and stride, which remains consistent even when the face is hidden.
+25% accuracy improvement in occluded environments.
Shoe & Body Analytics
Trained a specialized ResNet-50 model to generate 2048D embeddings for footwear and clothing color histograms. This "Appearance Vector" helps resolve identity when biometric traits are unavailable.
30% reduction in false positives (e.g., distinguishing two people wearing similar shirts by their shoes).
Low-Light Enhancement
Implemented a pre-processing layer using ZeroDCE++ and CLAHE to enhance low-light footage in real-time before feeding it into the recognition models.
+20% improvement in tracking stability during night shifts.
Business Impact
Tracking Accuracy
25% improvement in Re-ID accuracy, allowing security teams to reliably follow a suspect from the parking lot to the building interior without losing the track.
Operational Efficiency
35% reduction in manual review time, as operators no longer need to manually stitch together footage from different cameras to reconstruct a path.
False Alarm Reduction
30% decrease in false matches, significantly increasing operator trust in the automated alert system.
Relevant Enterprise Solutions
AI Search & Discovery Systems
AI Search & Discovery Systems Deliver more accurate, personalized, and intelligent search experiences with AI-powered semantic search and hybrid ranking....
View AI Search & Discovery SystemsRelated Services
Computer Vision
Computer Vision & Visual Intelligence Turn your cameras and documents into data sources. We build high precision Computer Vision Systems...
View Computer Vision