This paper proposes a sophisticated methodology for detecting pedestrian behaviors in nighttime surveillance videos in port environments. Given that ports operate continuously, reliable nighttime monitoring is critical, particularly under suboptimal lighting conditions. This research leverages advanced computer vision and deep learning techniques to enhance detection accuracy in low-visibility scenarios. Utilizing a dataset compiled from real port surveillance footage, we assess the efficacy of various deep learning models, particularly the integration of the SlowFast network with dark channel enhancement methods, such as dark channel prior and adaptive transmission refinement. These enhancements address the issues associated with low illumination. Our results demonstrate substantial improvements in detection accuracy compared to baseline approaches, thereby presenting a robust framework for augmenting safety and operational efficiency in ports.