Measure Localization Accuracy and Pipeline Performance #753

Establish a reproducible evaluation workflow for localization accuracy, execution time, end-to-end latency, and memory usage of the Awkernel preprocessing-to-NDT pipeline.

Scope:

• Evaluate position and orientation errors against timestamp-aligned ground-truth poses where available
• Measure execution time for crop box filtering, voxel grid downsampling, and NDT scan matching
• Measure end-to-end latency from scan availability at the pipeline input to publication of the corresponding localization result
• Distinguish processing time from queueing, scheduling, and data-transfer delays
• Measure peak memory usage, separating map preparation, persistent map storage, and per-scan processing
• Record workload characteristics, including point counts, NDT iteration counts, and convergence outcomes

Acceptance criteria:

• Metric definitions, timing boundaries, coordinate frames, and measurement methods are documented.
• Accuracy results include position and orientation error distributions, with ground-truth sources and alignment procedures specified.
• Datasets without ground truth are explicitly identified; reference-output differences are not reported as ground-truth accuracy.
• Timing results include median, tail percentiles, and maximum observed values, together with sample counts.
• Failed, non-converged, and dropped scans are reported rather than silently excluded.
• Peak memory measurements identify which allocations and buffers are included.
• Hardware, software versions, parameters, replay rate, warm-up procedure, and instrumentation overhead are recorded.
• Results are exported in a machine-readable format with reproducible evaluation commands.

Out of scope:

Formal worst-case execution-time guarantees, EKF integration, vehicle-control evaluation, and broad scheduler comparisons.