<p>In subway engineering applications, the detection of falling behaviors is susceptible to false positives and false negatives. Actions such as sitting, standing, bending over, and squatting are often misidentified as falls, leading to a significant number of misidentified detections. To enhance the reliability of engineering applications, it is necessary to improve and optimize existing models. In this paper, we propose a YOLO-CLIP algorithm model. Initially, YOLO is employed for the preliminary screening of falling targets, and the selected targets are then input into CLIP for a secondary judgment. By leveraging CLIP's rich text-image embedding capabilities, we filter out erroneous samples. Additionally, we propose an immune reinforcement learning algorithm for parameter optimization to identify the best parameters. Experiments conducted on both self-built and public datasets validate the effectiveness of the proposed algorithm, demonstrating a significant enhancement in model performance metrics.</p>