NP-YOLO: A Nighttime Pedestrian Detection Model for Video Surveillance

A target detection model, NP (Night Person)-YOLO, based on the YOLOv11s architecture, was proposed to address issues such as insufficient illumination, image noise interference, and low contrast in nighttime pedestrian detection. An adaptive enhancement module, RIE, based on Retinexformer, was first proposed. It enhanced input images to improve image quality and supported downstream detection tasks. Then, a convolutional additive token mixer and convolutional gated linear units were introduced into the key feature extraction module (C3k2 module). This significantly enhanced the model’s ability to extract target features in nighttime environments. Finally, the network structure of the detection head was optimized. This effectively reduced the model’s computational complexity while maintaining high detection accuracy.The experimental results show that NP-YOLO performs better than the YOLOv11s baseline model and other mainstream models on the NightSurveillance nighttime pedestrian dataset as well as on the ExDark and CityPersons public datasets. The [email protected] was improved by $2.2 \%$, $4.8 \%$, and $2 \%$. Furthermore, the model achieves a detection speed of 188.7 frames per second under the experimental conditions, meeting the realtime requirements of video surveillance detection.