Skip to main content
Back to timeline
arXivSource publication:

Road-user intention prediction moves to an unseen weather-and-time group, where the self-supervised objective holds 0.806 accuracy

Synopsis

Contrastive self-supervised pretraining for road-user intention prediction can retain environment-specific regularities when the objective is pooled over all training data, so nothing guarantees it holds up under precipitation and time-of-day combinations it never saw. CARE builds each window into an anchor, a positive view with late steps masked, and a negative view with that late segment perturbed, then adds an IRM-style gradient penalty on environment-wise contrastive risk so one shared head works across groups; masked reconstruction and evidence consistency are auxiliary constraints. On the IMPTC-derived four-class task the full objective reaches 0.840 accuracy and 0.023 expected calibration error under IID fine-tuning, and 0.806 on the selected held-out environment.

Source-provided article image: CARE: Contrastive Representation Learning with Environment-Invariance Regularization for Behavioural Intention Prediction
arXiv

Interpretation

Self-supervised pretraining for trajectory-based intention prediction now keeps contrastive risk low inside each recording environment, not only on average over the pooled batch. Representations learned with a pooled contrastive objective may retain environment-specific regularities, and neither standard supervised learning nor conventional self-supervised pretraining requires an encoder to avoid those environment shortcuts. After global fine-tuning on the IMPTC-derived four-class task, the full objective reaches 0.840 accuracy on the IID split against 0.830 for LiST and 0.821 for DLinear, and 0.806 on the selected held-out environment against 0.794 for DLinear and 0.785 for LiST.

Each input window becomes three views, with the first 60 steps identical across them and only the last 40 masked or perturbed, and two extra targets act on the masked steps: reconstructing the original motion features and keeping the anchor and masked views in agreement where input is visible. The positive view replaces contiguous late blocks with a learnable token and the negative view applies one of temporal roll, batch swap or time reversal to that same late segment; an InfoNCE loss pulls the anchor toward the positive and away from the negative, while an IRM-style penalty sums the squared gradients of each environment's contrastive risk with respect to a frozen reference head, standing in for the requirement that one shared head works in every group. The terms are trained jointly on windows cut from IMPTC trajectories at length 100 with 60 early and 40 later input steps and stride 80, giving 28272 training windows and 8155 test windows, with environments fixed by precipitation intensity and part of day.

Perspective

The results cover a four-class proxy intention task cut from IMPTC trajectories (straight, left, right, stop), with labels from kinematic rules and environments defined jointly by precipitation intensity and part of day; the held-out evaluation targets the single largest realised environment group, so the conclusions stay with that one shift. A next step the source sets out is to evaluate every adequately populated held-out group and to quantify multi-seed uncertainty.

The IRM-style penalty is a training objective and is not taken as evidence of formal invariance, and calibration on the held-out environment is not uniformly best: TCN reports ECE 0.039 against 0.090 for the full objective. That held-out result represents one fixed realised group, experiments use the fixed seed 2 without reporting seed-to-seed variability, the UMAP projections are qualitative and cannot establish independence between representation and environment labels, and the correspondence between numeric environment codes and the nominal precipitation-time categories is not fully documented in the available manuscript.