flutter_pose_detection 0.2.1

Hardware-accelerated pose detection using LiteRT (successor to TensorFlow Lite) with NPU/GPU/CPU acceleration. Detects 33 body landmarks in MediaPipe-compatible format.

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

0.2.1 - 2025-12-31

Fixed

• Android: Updated QNN SDK from 2.34.0 to 2.41.0
  • Fixes NoSuchMethodError: getBackendType()I crash on Snapdragon 8 Elite
  • Resolves native JNI method signature mismatch with newer Android firmware

Added

• Android: ChipsetDetector for runtime SoC detection (Snapdragon, Exynos, Tensor, MediaTek)
• iOS: CoreMLPoseDetector for native Core ML inference with ANE optimization
  • Direct MLModel API usage (no Vision framework overhead)
  • ImageNet normalization for HRNet models

0.2.0 - 2024-12-31

Changed

• BREAKING: Migrated from TensorFlow Lite to LiteRT 2.1.0 (Google's successor to TFLite)
• Android: New LiteRtPoseDetector using CompiledModel API with NPU/GPU/CPU fallback
  • Replaces TFLitePoseDetector and DelegateFactory
  • Uses Accelerator.NPU, Accelerator.GPU for hardware acceleration
  • Automatic fallback chain: NPU → GPU → CPU
• iOS: New LiteRtPoseDetector using TensorFlowLiteSwift 2.14.0
  • Replaces VisionPoseDetector with MoveNet-based detection
  • CoreML delegate for Neural Engine (A12+), Metal delegate for GPU
  • Automatic fallback chain: Neural Engine → Metal → CPU
• Both Platforms: Now use same MoveNet Lightning/Thunder models for consistent results
  • Cross-platform landmark positions match within 5% variance
  • Lightning (192x192) for fast mode, Thunder (256x256) for accurate mode

Removed

• TFLitePoseDetector.kt - replaced by LiteRtPoseDetector.kt
• DelegateFactory.kt - accelerator selection now handled by CompiledModel API
• VisionPoseDetector.swift - replaced by LiteRtPoseDetector.swift

Fixed

• Consistent 17-landmark COCO format output across platforms
• Proper resource cleanup on dispose()

0.1.0 - 2024-12-30

Added

• Initial release of NPU Pose Detection plugin
• iOS Support: Apple Vision Framework with Neural Engine acceleration
  • Automatic NPU/GPU acceleration on A12+ devices
  • VNDetectHumanBodyPoseRequest for 17 body landmarks
  • Landmark mapping to MediaPipe 33-point format
• Android Support: TensorFlow Lite with MoveNet Lightning model
  • GPU Delegate for hardware acceleration
  • NNAPI support for devices API 27-34
  • CPU fallback with XNNPack optimization
• Core Features:
  • Static image pose detection (detectPose, detectPoseFromFile)
  • Real-time camera frame processing (processFrame, startCameraDetection)
  • Video file analysis (analyzeVideo with progress tracking)
  • Configurable detection parameters (PoseDetectorConfig)
• Data Models:
  • Pose with 33 MediaPipe-compatible landmarks
  • PoseLandmark with normalized coordinates (0-1) and confidence scores
  • BoundingBox for detected person region
  • LandmarkType enum for easy landmark access
• Error Handling:
  • Typed DetectionError with error codes
  • Graceful fallback from NPU to GPU to CPU
• Example App:
  • Image detection demo with gallery picker
  • Real-time camera detection with pose overlay
  • Video analysis with progress UI