Topology-Guided Neural ODEs for Human Route Choice in Noisy 3D Scans

When a person walks through a cluttered room, the hard part of predicting their motion is often not the next step of their legs. It is which way around the table they will go. Current scene-aware motion predictors read the room as a cloud of 3D points and decode the future body frame by frame. That works when the scan is clean, but real scans have holes and jitter, and a hole can look like a doorway. We propose a motion model that keeps three things separate. The first is a learned description of how a body moves in continuous time, trained on large motion-capture data with no scene at all. The second is a short-range term that pushes the body away from nearby surfaces. The third is a route-choice term built from the persistent "holes" in walkable free space, meaning the obstacles you can walk around. These holes are tracked over time as the person's local view of the room changes, so that a short-lived artifact in the scan does not flip the predicted route.

The model is a neural ordinary differential equation, or neural ODE: a network that outputs the rate of change of a hidden state, which a numerical solver integrates to get the state at any future time (Neural Ordinary Differential Equations). The obstacle-holes come from zigzag persistence, a topology tool that follows loops and components through a sequence of shapes that can both grow and shrink. Our claim is narrow and testable. In the regime where route choice is ambiguous and scans are corrupted, a vector field guided by persistent holes predicts which side of each obstacle the person takes more accurately than a field guided by distance-to-surface alone. On clean scans with unambiguous routes, we expect no advantage, and we will report that honestly.

Core Claim

We write the hidden state as $h(t)$, which decodes to the body pose and the root (pelvis) position $x(t)$ on the floor. The dynamics are

$$\frac{dh}{dt} = f_{\text{kin}}(h) + g_{\text{near}}\big(h, \nabla d_S(x)\big) + g_{\text{route}}\big(h, \{w_k(x)\}_{k \in P_t}\big),$$

where $f_{\text{kin}}$ is a scene-free body-dynamics network and $d_S$ is the signed distance from $x$ to the scanned scene $S$. The set $P_t$ contains the free-space holes that are persistent at time $t$ under zigzag tracking of the person's local view. The quantity $w_k(x)$ is the accumulated winding angle of the root path around a representative cycle of hole $k$; it says which side of an obstacle the person is passing, and it is differentiable in $x$. If both scene terms are switched off, the model is a purely scene-free predictor. The environment-agnostic direction is therefore the base of the model and also a check on it, not something we discard. The new part is $g_{\text{route}}$: a continuous-time force that depends on topological features tracked over time, rather than on raw coordinates or a fixed snapshot.

Relation to Prior Work

Two papers supply the continuous-time machinery. Social ODE decodes multi-agent 2D trajectories with a latent ODE and adds hand-placed attractor and repeller points to the vector field at test time. It reports an 8.8% collision rate under sudden hazards versus over 22% for baselines. However, it treats agents as points and leaves dense static geometry unused. Our $g_{\text{near}}$ is the natural dense extension of its repellers, and we treat that extension as our strongest competitor, not as our contribution. Evaluation of Differentially Constrained Motion Models shows two things on vehicle data. Second-order neural ODEs beat rigid kinematic models (about 0.98 m versus 1.23–1.82 m ADE on rounD). And the Heun solver beats Euler by over 30%. This constrains our design: we use a lightly structured second-order field and a second-order solver, not hand-written body physics.

Two papers motivate the topological ingredient and also set its limits. Z-GCNETs feeds zigzag persistence images into a recurrent network and finds that loops carry more signal than clusters, with RMSE 31.61 versus 32.30 for AGCRN on PeMSD4. TEN-DM conditions a diffusion model on zigzag features of event images. Removing its topology module raised temporal error by 4.6%, and its topology-equipped model stayed robust under 5% coordinate noise. Both use topology as a summary vector fed to a discrete model, both predict mostly one step ahead, and neither involves physical agents. We use the cycle representatives themselves, extracted with methods like A Fast Algorithm for Computing Zigzag Representatives, as geometric objects inside the vector field. The snapshot-correspondence problem described by Rieck et al. is exactly why we use zigzag rather than per-frame persistence. Per-frame persistence cannot tell whether the hole ahead is the same table the person has been circling.

Scene-grounded motion work defines the task and the numbers to beat. GIMO fuses motion history, a PointNet++ scene encoding and gaze in a cross-modal transformer. GAP3DS turns gaze into affordance targets for a motion diffusion model. HUMANISE conditions a variational autoencoder on scene and language. All three decode in discrete steps over short horizons, and all three read the scene as metric point embeddings. At the other end, HVIS is a strong scene-free pose forecaster with no way to see obstacles. We do not claim topology replaces gaze or affordance cues. We claim it handles a specific failure that none of these models targets: choosing a route around obstacles when the geometry is noisy.

Research Program

The first stage trains $f_{\text{kin}}$ scene-free on AMASS and Human3.6M with a latent-ODE encoder–decoder and Heun integration. We check that it is competitive with scene-free forecasters on Human3.6M before any scene enters.

The second stage builds the topology pipeline. For each frame we project the scan to a 2D floor occupancy map and crop it to a visibility radius around the head. This mimics the egocentric partial view a deployed system would have. Consecutive crops are linked through their unions to form a zigzag sequence, and we compute 1-dimensional barcodes with cycle representatives. Holes whose zigzag lifetime falls below a threshold are dropped. That threshold is the one manual filtration choice, and we sweep it. Restricting the computation to 2D floor maps keeps cost manageable, which addresses the overhead both Z-GCNETs and TEN-DM report. For each surviving hole, $w_k$ is computed from its representative and fed, together with a learned per-hole embedding, to $g_{\text{route}}$. An auxiliary head predicts the final winding sign for each hole, trained against labels derived automatically from ground-truth trajectories.

The third stage trains both scene terms on GIMO and HUMANISE. During this stage $f_{\text{kin}}$ is frozen at first and then lightly fine-tuned. An optional gaze-target attractor, in the style of Social ODE, lets us test whether topology adds to gaze or only duplicates it.

Baselines

The naive reference point is constant root velocity with the last observed pose held fixed. A slightly stronger version uses our own $f_{\text{kin}}$ with both scene terms off. This second reference measures how much of every metric is achievable without seeing the room at all.

The strongest published scene-grounded methods are GIMO and GAP3DS, retrained on our splits from released settings where possible. GIMO reports 237.8 MPJPE with gaze and 262.1 without on known scenes, with destination translation error 0.245 versus 0.393 (GIMO). GAP3DS reports 0.95% penetration error versus 4.59% for SIF3D (GAP3DS). For route-choice accuracy and robustness to scan corruption, we found no reported result on Lacuna for any method, so we will measure all baselines ourselves. We do not expect to beat GAP3DS on penetration on clean scans. Our target there is to stay within its range while winning on route accuracy under corruption.

The decisive competitor is a controlled ablation: the same ODE, encoder, data, solver and tuning budget, but with $g_{\text{route}}$ removed and $g_{\text{near}}$ given a larger learned scene encoder (a PointNet++ embedding of the local crop) so its parameter count matches. This is "Social ODE with dense repellers plus a strong scene encoder". It is the model a reviewer would build instead. Every method gets the same hyperparameter search budget. The full model beats it only if, on the ambiguous-route subset and under corruption, route-choice accuracy improves by more than the seed-to-seed spread over five seeds, and penetration and MPJPE do not get worse.

Two further ablations isolate the topological ingredient. Replacing zigzag tracking with static persistence of each crop tests whether tracking over time matters. Feeding the same zigzag features to a discrete GRU decoder, in the style of Z-GCNETs, tests whether the continuous-time formulation matters. We also run an Euler-versus-Heun comparison, following Westny et al.

Why not just…?

Why not just give a strong point-cloud encoder the scan and let it learn route choice implicitly? This is the real threat, and the matched-capacity ablation above exists to settle it. The evidence that it may fall short is indirect. TEN-DM found that coordinate-based models degrade under 5% noise while topology-equipped ones do not, but that was on event data, not scans. If the encoder baseline matches us under hole and jitter corruption, we will conclude that topology is unnecessary for this task.

Why not just predict the destination and plan the shortest path to it? A planner fixes the route as whichever way around the obstacle is shortest. People often do not take that route, and nothing in the file measures how often. We will report the fraction of GIMO and HUMANISE trajectories whose route around each obstacle differs from the shortest path to the true endpoint. We will also run the planner with the ground-truth destination given. If that fraction is near zero, the planner wins and our route term is moot.

Why not just use a discrete transformer or diffusion decoder with topological features attached? The discrete-GRU ablation tests this directly. The case for continuous time rests on smoother long rollouts and on the ability to add field terms at test time, both demonstrated by Social ODE. It also allows querying sparse or irregular inputs. If the GRU variant matches us at 5-second horizons and under subsampled inputs, the ODE is not earning its solver cost.

Proposed Experiments

On GIMO known and unknown scenes, and on HUMANISE, we report MPJPE, destination error, penetration rate, goal distance and route-choice accuracy. The goal-distance reference is HUMANISE's reported 0.196 m for "lie down". Route-choice accuracy is the agreement of the predicted winding sign around each persistent obstacle with the ground truth, reported on all trajectories and on an ambiguous subset where both routes have similar length. We extend horizons beyond the usual 3–5 seconds where sequence length allows. We also subsample input frames irregularly, including a 2 fps gaze-like condition, to test continuous-time querying.

The decisive check is a corruption sweep on the scans. We vary point dropout, carve out synthetic holes that open false passages, and add coordinate jitter, all at increasing severity. Then we plot route-choice accuracy against severity for the full model, the matched point-encoder ODE, the static-persistence ablation and GIMO. The hypothesis is confirmed only if the zigzag model degrades measurably more slowly and the static-persistence ablation falls between the two. Two further checks round this out. Switching off the scene terms and evaluating on Human3.6M confirms that the scene-free core was not damaged by stage-three fine-tuning. Timing per predicted second confirms that the 2D zigzag computation stays within interactive budgets.