Loading...

Design and Implementation of an Adaptive Method for Efficient Pedestrian Detection in CCTV Cameras

Parvizi, Adel | 2024

0 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 57681 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Mohammadzadeh, Narjesolhoda; Shamsollahi, Mohammad Bagher
  7. Abstract:
  8. Pedestrian detection has been a significant area of interest both independently and as part of the object detection task. Classical algorithms always struggle with common environmental changes due to their simplistic approach to semantic feature extraction. In contrast, deep convolutional networks develop simple features through multiple layers, resulting in more reliable detections. Convolutional neural networks have demonstrated superior performance compared to classical algorithms, and they mostly run in real time thanks to graphic cards. However, speeding them up can reduce operational costs. This thesis focuses on static scenes and utilizes background information to make the detection process simpler. Under this assumption, moving objects are distinguishable in the scenes, and detecting pedestrians among them is straightforward. A classical algorithm is presented for detecting pedestrians in frames with few or no pedestrians. A modified version of YOLO object detector, GYOLO, will be introduced. The new network takes images as well as corresponding backgrounds as input that outperforms YOLO on COCO eval2017 data by 12.1% in recall and 10.3% in mAP50:95. GYOLO also outperforms YOLO on two videos of final three-video ultimate test data. We use the classic detector alongside the deep detectors to avoid unnecessary GPU usage. Therefore, two adaptable algorithms, AYOLO and AGYOLO, are introduced, the use of which, while reducing the use of the graphics processor, will not cause a big drop in evaluation criteria
  9. Keywords:
  10. Computer Vision ; Object Detection ; Convolutional Neural Network ; Pedestrian Detection ; Haar Pedestrian Detector ; Oriented Gradient Histogram ; Ip-Based Cameras

 Digital Object List

 Bookmark

No TOC