Loading...

High-Speed Human Tracking in CCTV Cameras

Mohammadi, Mohammad Javad | 2024

0 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 57979 (05)
  4. University: Sharif University of Technology
  5. Department: Electrical Engineering
  6. Advisor(s): Mohammadzadeh, Narjesolhoda
  7. Abstract:
  8. In this work, we present AppMoTrack, a novel tracking-by-detection (TBD) ap-proach that effectively balances speed and accuracy. The proposed framework lever-ages Incremental Principal Component Analysis (PCA) to extract robust features from bounding boxes, followed by Fisher Linear Discriminant Analysis (LDA) to refine these features by considering class separability, where each unique target across frames rep-resents a class. A hierarchical cost function is introduced to enhance data association precision. For detections with high confidence scores, a two-stage process is employed: first, a combination of Intersection over Union (IoU) and cosine distance between LDA-transformed PCA features; second, a comparison of HSV histograms for more nuanced matching. Detections with lower confidence scores are handled in a separate stage, ensuring robust association across varying detection qualities. Additionally, manual thresholds are replaced with an adaptive matching threshold mechanism based on K-means clustering, and track uncertainty is incorporated to improve decision-making during data association. A key innovation of AppMoTrack is its adaptive use of the detection network, dynamically determining when to run the detection model based on frame context. This reduces computational overhead by skipping frames adaptively. By integrating these methodologies, AppMoTrack achieves an optimal trade-off between ac-curacy and efficiency, advancing the state-of-the-art in reliable object tracking. On the MOT17 dataset, AppMoTrack achieves strong performance with the following metrics: HOTA: 63.21, MOTA: 79.39, IDF1: 77.12, and an FPS of 41.3 Hz
  9. Keywords:
  10. Human Tracking ; Adaptive Mechanism ; Ip-Based Cameras ; Tracking by Detection ; High Speed Camera ; Appearance Features

 Digital Object List

 Bookmark

...see more