Pedestrian alert system based on stereo distance prediction with object detection and segmentation

The pedestrian alert system is used to monitor specific areas, detect unauthorized or abnormal movements of pedestrians, and issue timely warnings. It is widely applied in critical fields such as power systems and border defense. Previous pedestrian distance measurement systems lacked corresponding datasets and were limited by the complexity of stereo algorithms. Additionally, background interference often caused measurement errors. To address these issues, we propose a pedestrian detection and stereo distance measurement-based warning system. First, a target detection model is used to detect pedestrian bounding boxes in the images. Considering the high real-time requirements of perimeter systems, we improved the coarse-to-fine Patch Match stereo algorithm to efficiently compute the distance of pedestrians within the bounding boxes. Meanwhile, we apply the Segment Anything Model segmentation model to accurately segment the disparity map within the pedestrian bounding boxes, removing background interference and occlusions, significantly improving the accuracy of distance calculation. With this approach, we provide a precise pedestrian warning capability for the surveillance system.