VioEstimator API

VioEstimator<T, Backend = MsckfBackend<T>> (sdk/include/branes/sdk/vio_estimator.hpp, #45) is the public façade the daemons drive. It owns the visual front end and the selected estimator backend, and exposes a small, lifecycle-gated API.

The API

void configure(const VioConfig&);              // fix parameters, init the backend
void activate();   void deactivate();          // managed lifecycle
void teardown();
void feed_imu(std::span<const ImuMeasurement<T>>);     // inputs by view
void feed_image(double t, cv::Image<const std::uint8_t>);
Pose      current_pose()  const;               // SE3 T_world_imu
NavState  current_state() const;
void      reset();

Inputs are non-owning views — an Image view for frames, a span for IMU batches; the estimator copies nothing it doesn’t need to retain.

Lifecycle — mirroring the Rust RM

The estimator’s states mirror the Resource Manager’s managed lifecycle (Resource Manager):

Unconfigured ──configure──▶ Inactive ──activate──▶ Active
                              ▲                       │
                              └────deactivate/reset───┘   teardown ─▶ (terminal)

Measurements are consumed only while Active — feed_imu/feed_image are no-ops otherwise. This honors the “no dynamic reconfiguration on the hot path” rule: parameters are fixed at configure.

What feed_image does

It runs the front-end loop (pyramid → KLT-track existing features → re-detect FAST corners, suppressing near existing tracks → assign stable ids), producing one FrontendObservation per surviving track, and hands them to the backend’s process_camera. A telemetry accessor, num_tracked_features(), exposes the current track count.

Backend substitutability

Because Backend is a template parameter constrained only by the VioBackend interface, the estimator can be instantiated on MsckfBackend<T> (the default) or on the SlidingWindowBackend<T> skeleton — the front end and the public API are identical either way.

Validated

Tests cover the lifecycle gating (feeds ignored unless Active; correct transitions; reset/teardown) and a full front-end → backend run on translating synthetic frames that detects and persistently tracks features, keeps the covariance PSD and the clone window bounded, and produces a finite pose. The runnable EuRoC replay example lives in the dataset harness (#46).