Scholarlink

<p>Detection of pedestrians accurately is a critical requirement for various video surveillance applications. The existing techniques still face limitations in precisely identifying a resembling non-human object (like a mannequin) among pedestrians, leading to high false positives and eventually low accuracy. To address this issue, we propose a novel system for pedestrian detection by filtering non-pedestrian detections using an automatically segmented walkable area. Our contributions include (a) an algorithm to automatically detect seed points for the Segment Anything Model using optical flow and a pre-trained YOLOv3 model, (b) an algorithm for segmenting walkable area from the captured surveillance video, and (c) an algorithm for pedestrian detection using fine-tuned YOLOv3 tiny and detected walkable area. The pretrained YOLOv3 tiny is fine-tuned for the person class using the COCO dataset and tested on the Oxford Town Centre video surveillance dataset. Our integrated solution of pedestrian detection (SWAY) shows an improved AP50 over standalone YOLOv3 tiny on the Oxford Town Center video surveillance dataset and achieves almost the same FPS. Also, there is a decrease of 0.58 in FPPI using SWAY in comparison to the YOLOv3 tiny model.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Sway: efficient pedestrian detection using SAM-based walkable area segmentation and YOLO

  • 期刊论文

  • 发表日期：

    2025年7月2日

查看全文
  • Jagrati Gupta,
  • Shobha Sharma

摘要

Detection of pedestrians accurately is a critical requirement for various video surveillance applications. The existing techniques still face limitations in precisely identifying a resembling non-human object (like a mannequin) among pedestrians, leading to high false positives and eventually low accuracy. To address this issue, we propose a novel system for pedestrian detection by filtering non-pedestrian detections using an automatically segmented walkable area. Our contributions include (a) an algorithm to automatically detect seed points for the Segment Anything Model using optical flow and a pre-trained YOLOv3 model, (b) an algorithm for segmenting walkable area from the captured surveillance video, and (c) an algorithm for pedestrian detection using fine-tuned YOLOv3 tiny and detected walkable area. The pretrained YOLOv3 tiny is fine-tuned for the person class using the COCO dataset and tested on the Oxford Town Centre video surveillance dataset. Our integrated solution of pedestrian detection (SWAY) shows an improved AP50 over standalone YOLOv3 tiny on the Oxford Town Center video surveillance dataset and achieves almost the same FPS. Also, there is a decrease of 0.58 in FPPI using SWAY in comparison to the YOLOv3 tiny model.