To alleviate the accumulation of risk losses caused by delayed detection of abnormal events in traffic surveillance videos under constrained computing capacity, a non-homogeneous Poisson process was constructed to characterize the arrival process of abnormal events. An increasing convex delay-loss function grounded in survival analysis was then introduced to investigate the risk evolution of abnormal events, and a risk monitoring and analysis framework for abnormal events in traffic surveillance videos, namely “event arrival–delayed detection–loss accumulation,” was established. In addition, a dynamic polling optimization model was developed to generate differentiated polling schedules for large-scale traffic video scenarios, thereby yielding an optimization scheme that balances risk control and computing-resource allocation. The results show that the proposed framework can jointly incorporate computing resources, polling cadence, and risk growth within a unified objective function; the hourly dynamic optimization scheme reduces the cumulative risk loss from 8 041.12 to 7 201.29, achieving an overall improvement rate of 10.44% for the full sample and still yielding an improvement rate of 8.94% in sparse-event small-sample scenarios; under parameter perturbations of the loss function, the model still maintains a stable positive improvement effect. These findings indicate that the proposed optimization model can effectively reduce the cumulative risk loss associated with abnormal road events without additional computing resources, while maintaining satisfactory robustness and scheduling fairness.