Related Experiment Video
Updated: May 4, 2026
08:25
Combining Eye-tracking Data with an Analysis of Video Content from Free-viewing a Video of a Walk in an Urban Park Environment
Published on: May 7, 2019
8.6K
Scene-specific pedestrian detection for static video surveillance.
Xiaogang Wang1, Meng Wang2, Wei Li2
1The Chinese University of Hong Kong, Hong Kong.
IEEE Transactions on Pattern Analysis and Machine Intelligence
|December 21, 2013
Summary
This study introduces an automated method to adapt generic pedestrian detectors for specific scenes in video surveillance. The approach significantly enhances detection accuracy without manual labeling, improving performance by up to 48%.
Area of Science:
- Computer Vision
- Machine Learning
- Artificial Intelligence
Background:
- Generic pedestrian detectors often fail in specific environments due to data distribution mismatches.
- Manual annotation of target scene data is time-consuming and resource-intensive for adapting detectors.
Purpose of the Study:
- To develop an automated transfer learning framework for adapting generic pedestrian detectors to scene-specific applications in static video surveillance.
- To eliminate the need for manual labeling of target scene samples during the adaptation process.
Main Methods:
- A four-step transfer learning framework utilizing visual affinity graphs and context cues.
- Source sample weighting based on target sample distribution and confidence score propagation.
- A novel confidence-encoded Support Vector Machine (SVM) objective function to guide detector training.
Main Results:
- Significant improvements in pedestrian detection rates: 48% and 36% at one false positive per image (FPPI) on two datasets.
- The training process demonstrated rapid convergence, typically within one or two iterations.
- The scene-specific detector, when tested, relied solely on appearance-based features without context cues.
Conclusions:
- The proposed automated transfer learning method effectively adapts generic pedestrian detectors to specific surveillance scenes.
- The confidence-encoded SVM approach offers a robust alternative to hard thresholding, enhancing transfer learning performance.
- This technique provides a practical solution for improving pedestrian detection accuracy in diverse, real-world video surveillance scenarios.
