Autonomous collision avoidance for electric vertical takeoff and landing (eVTOL) aircraft in structured urban low-altitude airspace must satisfy both physical collision-prevention requirements and operational rules. This study proposes a rule-constrained dynamic velocity vector optimization (RC-DVVO) framework. First, urban low-altitude operational rules, including no-fly zones, flight corridors, altitude layers, and speed limits, are decomposed and mapped into a time-varying rule-feasible domain in velocity space. Second, this domain and the kinematically executable domain are incorporated into a three-dimensional dynamic velocity obstacle model, with collision probability and time to collision used to construct a risk term. Neighboring aircraft intentions and ground-recommended velocities are then integrated into a multi-objective cooperative optimization model, which is solved in real time using rule-based pruning, risk prediction, and local continuous optimization. Simulations of multiple crossing conflicts in an urban canyon show that RC-DVVO achieves a collision-avoidance success rate of 99%, a rule-violation rate of 0.8%, a mean curvature of 0.0132 rad/m, and a mean decision time of 12.3 ms. Compared with classical DVVO, the success rate increases by 5 percentage points, the rule-violation rate decreases by 88.9%, and the mean curvature decreases by 28.6%, while the decision time meets the requirements of 20 Hz control. These results indicate that the method enables safe and smooth autonomous eVTOL collision avoidance within modeled compliance boundaries and provides a basis for incorporating computable rules into collision-avoidance decisions in low-altitude intelligent networks.

This study develops an eight-indicator framework for describing low-altitude operational risk and a candidate air-ground response architecture. The evidence is drawn from three distinct supplementary datasets: 1,000 source-labelled sortie records (772 retained after exposure and duration quality control; 740 operational and 32 simulated), a trajectory/alert slice aggregated to 123 occupied 500 m × 500 m by 5-min cells, and 6,079 occupied 5-min windows from 55 days of operational records. These datasets are analysed separately. In the 55-day series, the adjusted negative-binomial model estimated a rate ratio of 3.32 per one-level increase in mean sortie risk score (95% CI 3.13–3.52), while density had an independent rate ratio of 1.11 (95% CI 1.07–1.14). In the 700-cell trajectory pipeline, the score–alert association was not statistically distinguishable from unity (RR 1.0026 per point, 95% CI 0.9957–1.0095; p = 0.464). Near-perfect rolling-window AUCs were attributable to the close constructional dependence between sortie risk scores and event records and are not interpreted as independent predictive validation. Because the observations are low-altitude UAV operations and simulation records rather than commercial passenger eVTOL flights, the findings support a pilot assessment workflow, not eVTOL accident-risk estimates or validated operational thresholds.