CreaVision

Technical Scope

Multi-camera surveillance and monitoring environments fail when tracking models lose subject identity during occlusions or blind-spot handoffs.

CreaVision established an in-house visual re-identification pipeline that extracts identity embeddings and associates them across disparate camera streams.

Architecture Highlights

• Spatial-Temporal Re-ID: Extracted visual feature vectors coupled with temporal movement heuristics to disambiguate targets re-entering scenes.
• Optimized Detection & Association: Integrated pruned Mask R-CNN feature extractors with an optimized DeepSORT association matrix, achieving real-time multi-camera frame processing on edge hardware.