Release Notes: Behavioral Analysis Service#

Version 1.0.0#

Release date: September 9, 2026

Summary:

Initial release of the Behavioral Analysis Service, a pose-based suspicious activity detection microservice for retail loss-prevention use cases. This is a production-ready microservice that leverages OpenVINO Runtime for efficient edge inference without requiring PyTorch dependencies.

The Behavioral Analysis Service analyzes video frame sequences to detect suspicious behaviors by:

  1. Extracting skeletal pose keypoints using YOLO26n-pose optimized with OpenVINO Runtime.

  2. Evaluating pose sequences against behavioral patterns defined in YAML.

  3. Optionally forwarding key frames to a Visual Language Model (VLM) for visual verification.

Features:

  • Pose extraction — YOLO26n-pose inference via OpenVINO Runtime (no PyTorch dependency)

  • Declarative YAML behavioral pattern engine — add new patterns without code changes

  • Built-in shelf_to_waist concealment detection pattern — targeting retail shrinkage scenarios

  • Optional VLM confirmation — via OpenVINO Model Server (Qwen2.5-VL-7B-Instruct)

  • Event-driven MQTT processing — ba/requests → ba/results with entity deduplication

  • Async SeaweedFS (S3-compatible) frame retrieval — via aioboto3

  • Circuit breaker in VLM client — 3-failure threshold, 30-second cooldown

  • Entity deduplication and max-concurrency backpressure — in the MQTT consumer

  • Base image — intel/dlstreamer:2026.2.0-ubuntu24 (Python 3.12)

  • Container-ready — fully configurable via environment variables and volume-mounted YAML

Use Cases:

  • Retail loss prevention — detect suspicious concealment behaviors (e.g., shelf-to-waist movements) in real time

  • Behavioral analysis at the edge — extract and evaluate pose sequences without reliance on cloud inference

  • Multimodal detection — combine pose-based detection with optional VLM visual verification for improved accuracy

  • Video surveillance — efficient frame-by-frame behavioral monitoring in retail environments

Known Limitations:

  • The service requires a reachable SceneScape deployment (MQTT broker + SeaweedFS) to produce meaningful output

  • VLM confirmation adds latency; consider circuit breaker settings for high-throughput scenarios

  • YOLO26n-pose inference performance is hardware-dependent; refer to System Requirements for supported compute devices

  • Pattern definitions are YAML-based and require validation before deployment