NeuroFlow: A Non-Linear Dual-Channel Fusion Framework for Adaptive Traffic Signal Control with Siren-Centric Long-Range Emergency Vehicle Preemption

Adaptive traffic signal control (ATSC) in heterogeneous, low-compliance urban road environments
presents challenges that existing systems address inadequately. Static background-subtraction methods
lose accuracy under illumination change; single-channel object detectors fail at night; body-based
emergency vehicle preemption cannot operate where ambulance livery is non-standardised. This paper
presents NeuroFlow, a three-tier intelligent traffic management framework designed for precisely these
conditions. The first contribution is the Illumination-Adaptive Photometric Golden-Reference (IAPGR)
method, which selects from a 12-image photometric calibration bank and applies closed-form brightness
normalisation before pixel-difference occupancy computation, eliminating the need for per-intersection
model retraining. The second contribution is the Normalised Product Fusion (NPF) operator, a provably
unbiased non-linear density fusion formula that combines IAPGR occupancy (Channel A) and YOLOv5
vehicle-count density (Channel B) through a weighted harmonic structure. NPF reduces density mean
absolute error (MAE) by 38.8% relative to YOLO-alone and by 46.4% relative to IAPGR-alone, as
confirmed by a three-configuration ablation study. The third contribution is the Siren-Centric Emergency
Vehicle Preemption Subsystem (SCEVPS), which detects emergency light flash signatures at 600 m
range — independent of vehicle body classification — providing a 36-second proactive corridor
drainage window. Vehicle and siren detectors, both YOLOv5lu, achieve mAP@50 = 0.794 and
mAP@50 = 0.750 on heterogeneous mixed-traffic validation sets. Two complementary validation
modalities are reported: (i) a SUMO two-scenario microscopic simulation against a fixed-time baseline,
demonstrating a 96.4% reduction in mean intersection waiting time (29.48 s to 1.06 s), 81.3% reduction
in mean time-loss, complete elimination of ambulance and police signal waiting time, and 36.0%
reduction in mean trip duration; and (ii) offline testing on recorded Pakistani CCTV intersection footage,
confirming IAPGR stability through a sudden weather transition. The system outperforms SCOOT by
84.4 percentage points and deep reinforcement learning baselines by 74.4 percentage points in
waiting-time reduction under equivalent demand conditions. Limitations, including siren detector recall
(0.66) and single-intersection simulation scope, are discussed with planned mitigations.