svo-cpp 1.5.0

Semi Direct Visual Odometry with python bindings

SVO CPP

Semi Direct Visual Odometry with Python bindings.

Installation

pip install svo-cpp

TODO (Click me)

• Implement Full MapGraph Creation

  • The current get_map_graph() in SVOEngine is a placeholder that only returns the last frame.
  • Sub-task: Add C++ pybind11 functions to iterate through all active keyframes in the svo::Map.
  • Sub-task: Convert the C++ keyframe data into a list of Python MapNode objects to provide a complete map representation to the SLAMService.

• Resolve Feature Handling Mismatch

  • The Python VOEngine interface provides pre-computed features to process_frame, but the C++ SVO library performs its own feature detection and ignores them.
  • Decision: Choose a long-term strategy:
    1. (Recommended for SVO): Keep the current implementation and clearly document that SVO handles its own feature detection.
    2. (Advanced): Modify the C++ FrameHandlerMono to accept and use external features, bypassing its internal FAST detector. This would allow for experimentation with different feature detectors from Python.

• Finalize ARM-Specific Parameter Tuning

  • Methodically test and validate a final jetson_config dictionary with optimal parameters for the drone's hardware and expected motion patterns.
  • Sub-task: Focus on finding the best balance for reproj_thresh, as it's the most critical parameter for tracking stability.
  • Sub-task: Tune init_min_disparity to ensure reliable initialization in real-world drone startup scenarios (e.g., slow takeoff).

• Investigate Compiler Flag Impact

  • The performance difference between x86 and ARM suggests sensitivity to compiler optimizations.
  • Sub-task: Compile and test the C++ modules using the -O2 optimization level instead of -O3 to see if it improves numerical stability.
  • Sub-task: Double-check all CMakeLists.txt files to ensure the -ffast-math flag (which can reduce precision) is not being used.

• Integrate IMU Data for VIO

  • The SLAMService is already designed to handle IMU data for visual-inertial odometry. The SVO library also has capabilities for this.
  • Sub-task: Create C++ bindings for SVO's IMU processing functions.
  • Sub-task: Implement the logic in SVOEngine to pass IMU data from the _on_calib_sync callback to the C++ backend.

• Enable Dynamic Feature Filtering

  • Your application can provide a dynamic mask to filter features on moving objects (e.g., other drones, people).
  • Sub-task: Modify the C++ FrameHandlerMono to accept an image mask.
  • Sub-task: Apply this mask during the internal feature detection step to ignore features in dynamic regions.

• Improve Relocalization Logic

  • The logs show the system enters a RELOCALIZING state frequently on ARM.
  • Sub-task: Expose C++ parameters related to relocalization to the Python set_svo_config function.
  • Sub-task: Tune these parameters to make relocalization faster and more reliable.

• Refine State Management

  • The mapping from the C++ Stage enum to the Python SLAMState enum is functional but could be more detailed to provide better system health information.
  • Sub-task: Provide more granular state updates from the C++ backend to the Python SVOEngine.