flutter_pose_detection 0.4.1

Hardware-accelerated pose detection using MediaPipe PoseLandmarker with GPU/NPU acceleration. Detects 33 body landmarks.

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.

0.4.1 - 2026-01-02

Fixed

• Fixed incorrect model name in Platform Support table (movenet_lightning → pose_landmarks_detector)

0.4.0 - 2026-01-02

Added

• NPU Support (Android): Qualcomm QNN delegate for Snapdragon devices
  • Battery-efficient inference on Hexagon DSP (HTP backend)
  • ~13-16ms inference time with significantly lower power consumption
  • Automatic skel library extraction from bundled assets
  • Supports Snapdragon 8 series (V68-V79 HTP versions)
• New preferredAcceleration config option for explicit backend selection
  • AccelerationMode.npu - Force NPU (Snapdragon only)
  • AccelerationMode.gpu - Force GPU (default behavior)
  • null - Auto-select best available
• Backend switching between MediaPipe (GPU) and TFLite+QNN (NPU)
• Example app with NPU toggle switch and benchmark functionality

Changed

• NpuPoseDetector now supports runtime backend selection
• Updated README with NPU documentation and performance comparison table

0.3.3 - 2026-01-02

Fixed

• Fixed Dart formatter issues for pub.dev static analysis
• Updated plugin_platform_interface constraint (^2.0.2 → ^2.1.0) for downgrade compatibility

0.3.2 - 2026-01-02

Fixed

• Fixed Android minSdkVersion in README (24 → 31)
• Separated Model Architecture section for iOS and Android

0.3.1 - 2026-01-02

Changed

• Updated README with comprehensive API documentation
• Added Configuration, Camera Stream, Video Analysis, and Angle Calculation sections
• Fixed Quick Start code examples (MediaPipeLandmarkType → LandmarkType)
• Added API Reference table with all methods and properties
• Improved MediaPipe 33 Landmarks documentation with structured table

Fixed

• Corrected library documentation (Vision Framework → MediaPipe PoseLandmarker)
• Fixed Pose class comment (17 landmarks → 33 landmarks)

0.3.0 - 2026-01-01

Changed

• BREAKING: Migrated from LiteRT/MoveNet to MediaPipe PoseLandmarker API
• Android: Uses official MediaPipe Tasks Vision library with GPU delegate
  • pose_landmarker_lite.task model for 33-landmark detection
  • Automatic GPU → CPU fallback
• iOS: Uses TFLite with CoreML/Metal delegates for ANE/GPU acceleration
  • pose_detector.tflite + pose_landmarks_detector.tflite 2-stage pipeline
  • Automatic ANE → GPU → CPU fallback

Removed

• LiteRT/MoveNet fallback code (simplified architecture)
• LiteRtPoseDetector.swift - MediaPipe only
• movenet_lightning.tflite, movenet_thunder.tflite - replaced by MediaPipe models
• hrnetpose_w8a8.tflite, HRNetPose.mlpackage - no longer needed

Fixed

• iOS main thread blocking during initialization (moved to background thread)
• iOS coordinate normalization (pixel space 0-256 → normalized 0-1)
• Android model loading with ByteBuffer instead of file path

0.2.2 - 2025-12-31

Fixed

• Critical: Fixed pose coordinate mismatch with video aspect ratio
  • Added letterboxing (padding) during image preprocessing to maintain original aspect ratio
  • Previously, images were stretched to 192x256 input size, causing coordinate distortion
  • Coordinates now correctly map back to original image dimensions
• Android: Implemented letterboxBitmap() and transformKeypointsFromLetterbox() in LiteRtPoseDetector.kt
• iOS: Implemented letterboxAndExtractRGB() and transformKeypointsFromLetterbox() in CoreMLPoseDetector.swift

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