Uncertainty-Aware Counterfactual Traffic Signal Control with Predictive Safety and Starvation-Avoidance Constraints Using Vision-Based Sensing

TL;DR

This paper tackles vision-based traffic signal control under partial observability by casting the problem as a POMDP and solving it with a belief-space constrained MPC (CS-MPC) approach called UCATSC. It introduces expert-designed hard constraints for predictive safety (dilemma-zone risk) and fairness (starvation barriers), along with counterfactual belief-space rollouts and a movement-level belief representation to enable real-time operation. The method achieves greater robustness to perception uncertainty, reduces extreme safety-risk events, and lowers emission proxies compared with queue-based and some baselines, while maintaining practical runtimes (~$110$--$120$ ms per step). UCATSC provides interpretable, model-based control with explicit safety guarantees, offering a scalable path toward deployable vision-based TSC that balances efficiency, safety, and fairness in real-world settings.

Abstract

Real-world deployment of adaptive traffic signal control, to date, remains limited due to the uncertainty associated with vision-based perception, implicit safety, and non-interpretable control policies learned and validated mainly in simulation. In this paper, we introduce UCATSC, a model-based traffic signal control system that models traffic signal control at an intersection using a stochastic decision process with constraints and under partial observability, taking into account the uncertainty associated with vision-based perception. Unlike reinforcement learning methods that learn to predict safety using reward shaping, UCATSC predicts and enforces hard constraints related to safety and starvation prevention during counterfactual rollouts in belief space. The system is designed to improve traffic delay and emission while preventing safety-critical errors and providing interpretable control policy outputs based on explicit models.

Figures (7)

• Figure 1: Vision-based traffic signal control under partial observability. The true microscopic traffic state $x_t$ evolves via the dynamics model and is observed through a noisy vision measurement $z_t$. UCATSC maintains a structured movement-level belief $b_t=\{p(Q_{m,t}),p(\lambda_{m,t}),\tau_{m,t}\}_{m\in\mathcal{M}}$, and selects a signal action $u_t$ via constrained counterfactual rollouts (CS-MPC) subject to predictive safety and starvation-avoidance constraints.
• Figure 2: Mean detection probability $p_{\mathrm{det}}$ vs. time (baselines, 30 s smoothing). UCATSC maintains higher and more stable detection confidence during vision degradation.
• Figure 3: Occlusion proxy $(1-p_{\mathrm{det}})$ vs. time (baselines, 30 s smoothing). UCATSC suppresses worst-case occlusion spikes relative to fixed-time control.
• Figure 4: Risk proxy distribution across baselines (900 s window). UCATSC reduces extreme tail risk while maintaining a comparable median.
• Figure 5: Total emission proxy distribution across baselines (per-run). UCATSC achieves substantially lower emissions than queue-based control.