Red light crossing violations (RLCV) pose a significant hazard due to various factors influencing driver behavior and traffic signal operations. This study explores the efficacy of Deep Residual Neural Networks (DRNNs) in traffic signal optimization, specifically examining their influence on RLCV frequency. Data was collected from twenty signalized intersections over fifteen-minute intervals during weekdays, focusing on traffic volume, signal timing, geometric characteristics of approaches, and instances of the RLCV. The model successfully managed the well-known deep learning problem of vanishing gradient by exploiting DRNNs’ complex design, distinguished by their residual learning framework and identity mapping, easing the training of extremely deep networks. This enables the precise prediction and measurement of traffic flow and RLCV under changing circumstances, with R² = 0.9 for the testing data. The proposed methodology, which requires up to 48,000 samples, guarantees that the variance-based sensitivity analysis method’s indices converge, offering solid insights into the system’s behavior. The findings showed that the maximum queue length (QLmax) and the green and cycle times (G and C) substantially influence RLCV frequency, with a noticeable rise when QLmax exceeds ten vehicles and the G/C ratio falls below 0.15. This study underlines the urgency of addressing these factors to reduce RLCV frequency. It also emphasizes the potential of DRNNs in traffic management and recommends that future research concentrate on integrating real-time data for dynamic traffic signal modifications, therefore maximizing DRNNs’ potential in this sector.