/* fuse.js — fuseWorld: combine per-marker camera pose estimates into one world pose. * * Each detected marker yields a camera-in-world estimate: * cameraWorld = anchorWorld * inverse(markerPoseInCamera) * Estimates are fused by confidence-weighted quaternion slerp (incremental * weighted average) and weighted position mean. Confidence = marker screen area * (bigger/closer markers dominate). A light temporal smooth removes residual jitter. * * Requiring >= 2 visible markers is handled upstream by anchor placement density; * fuseWorld itself works with 1..N. */ import * as THREE from 'three'; import { anchorWorldMatrix } from './pose.js'; const _inv = new THREE.Matrix4(); const _m = new THREE.Matrix4(); const _p = new THREE.Vector3(); const _q = new THREE.Quaternion(); const _s = new THREE.Vector3(); export class WorldFuser { constructor() { this.anchorMats = new Map(); // markerId -> Matrix4 this.smoothPos = null; this.smoothQuat = null; this.posAlpha = 0.35; // smoothing factors (higher = snappier) this.quatAlpha = 0.35; this.lastFuseT = 0; } setScene(scene) { this.anchorMats.clear(); for (const a of scene.anchors || []) { if (a.enabled === false) continue; this.anchorMats.set(a.markerId, { mat: anchorWorldMatrix(a), sizeMM: a.sizeMM || 60 }); } } sizeFor(markerId) { return this.anchorMats.get(markerId)?.sizeMM || 60; } knownIds() { return new Set(this.anchorMats.keys()); } /** estimates: [{ markerId, position(mm), quaternion, area }] */ fuse(estimates) { const MM = 0.001; const cams = []; for (const e of estimates) { const entry = this.anchorMats.get(e.markerId); if (!entry) continue; // marker pose in camera space -> matrix (translate mm->m) _m.compose(_p.copy(e.position).multiplyScalar(MM), e.quaternion, _s.set(1, 1, 1)); _inv.copy(_m).invert(); // camera in marker space const camWorld = new THREE.Matrix4().multiplyMatrices(entry.mat, _inv); const pos = new THREE.Vector3(); const quat = new THREE.Quaternion(); camWorld.decompose(pos, quat, _s); cams.push({ pos, quat, w: Math.max(1, e.area) }); } if (!cams.length) return null; // weighted position mean + incremental weighted slerp for orientation let wSum = cams[0].w; const pos = cams[0].pos.clone().multiplyScalar(cams[0].w); const quat = cams[0].quat.clone(); for (let i = 1; i < cams.length; i++) { const c = cams[i]; // hemisphere alignment before slerp (quaternion double-cover) if (quat.dot(c.quat) < 0) c.quat.set(-c.quat.x, -c.quat.y, -c.quat.z, -c.quat.w); const t = c.w / (wSum + c.w); quat.slerp(c.quat, t); pos.add(c.pos.clone().multiplyScalar(c.w)); wSum += c.w; } pos.multiplyScalar(1 / wSum); // temporal smoothing const now = performance.now(); if (this.smoothPos && now - this.lastFuseT < 500) { this.smoothPos.lerp(pos, this.posAlpha); if (this.smoothQuat.dot(quat) < 0) quat.set(-quat.x, -quat.y, -quat.z, -quat.w); this.smoothQuat.slerp(quat, this.quatAlpha); } else { this.smoothPos = pos.clone(); this.smoothQuat = quat.clone(); } this.lastFuseT = now; return { position: this.smoothPos.clone(), quaternion: this.smoothQuat.clone(), markerCount: cams.length }; } }