Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Low-Latency Video Anonymization for Crowd Anomaly Detection: Privacy vs. Performance

Mulugeta Weldezgina Asres, Lei Jiao, Christian Walter Omlin

TL;DR

A novel lightweight adaptive anonymization for VAD (LA3D) that employs dynamic adjustment to enhance privacy protection and demonstrates that LA3D enables substantial improvement in the privacy anonymization capability without majorly degrading VAD efficacy.

Abstract

Recent advancements in artificial intelligence promise ample potential in monitoring applications with surveillance cameras. However, concerns about privacy and model bias have made it challenging to utilize them in public. Although de-identification approaches have been proposed in the literature, aiming to achieve a certain level of anonymization, most of them employ deep learning models that are computationally demanding for real-time edge deployment. In this study, we revisit conventional anonymization solutions for privacy protection and real-time video anomaly detection (VAD) applications. We propose a novel lightweight adaptive anonymization for VAD (LA3D) that employs dynamic adjustment to enhance privacy protection. We evaluated the approaches on publicly available privacy and VAD data sets to examine the strengths and weaknesses of the different anonymization techniques and highlight the promising efficacy of our approach. Our experiment demonstrates that LA3D enables substantial improvement in the privacy anonymization capability without majorly degrading VAD efficacy.

Low-Latency Video Anonymization for Crowd Anomaly Detection: Privacy vs. Performance

TL;DR

A novel lightweight adaptive anonymization for VAD (LA3D) that employs dynamic adjustment to enhance privacy protection and demonstrates that LA3D enables substantial improvement in the privacy anonymization capability without majorly degrading VAD efficacy.

Abstract

Recent advancements in artificial intelligence promise ample potential in monitoring applications with surveillance cameras. However, concerns about privacy and model bias have made it challenging to utilize them in public. Although de-identification approaches have been proposed in the literature, aiming to achieve a certain level of anonymization, most of them employ deep learning models that are computationally demanding for real-time edge deployment. In this study, we revisit conventional anonymization solutions for privacy protection and real-time video anomaly detection (VAD) applications. We propose a novel lightweight adaptive anonymization for VAD (LA3D) that employs dynamic adjustment to enhance privacy protection. We evaluated the approaches on publicly available privacy and VAD data sets to examine the strengths and weaknesses of the different anonymization techniques and highlight the promising efficacy of our approach. Our experiment demonstrates that LA3D enables substantial improvement in the privacy anonymization capability without majorly degrading VAD efficacy.

Paper Structure

This paper contains 14 sections, 8 equations, 8 figures, 9 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Literature Review
  3. Methodology
  4. Problem Formulation
  5. Lightweight Adaptive Full-Body Anonymization
  6. Privacy Attribute Classification Model
  7. Video Anomaly Detection Model
  8. Results and Discussion
  9. Evaluation Approach and Performance Metrics
  10. Privacy Leakage Detection
  11. Privacy Attribute Detection (PD)
  12. Person Re-Identification (ReID)
  13. Video Anomaly Detection
  14. Conclusion

Figures (8)

  • Figure 1: Conventional AN techniques on sample images from the VISPR dataset: (left to right) RAW_IMAGE, BLACKENED, PIXELIZED, and BLURRED. Blurring and pixelization with fixed AN parameters are not optimized to handle the target figures in images that differ in size or are positioned at varying depths.
  • Figure 2: The privacy attribute class distribution of the VISPR train set. The estimated relative label distribution is $D_c~=~[51.53\%, 46.25\%, 44.18\%, 5.88\%, 47.1\%, 9.64\%]$ for the No_Attribute, Gender, Face, Nudity, Color, and Relationship attributes, respectively.
  • Figure 3: Visual comparison of different AN methods on sample images from the VISPR dataset: (left to right) 1:RAW_IMAGE, 2:BLACKENED, 3:BLACKENED_EDGED, 4:PIXELIZED_D2, 5:PIXELIZED_D4, 6:PIXELIZED_D8, 7:BLURRED, 8:PIXELIZED_D2_A ($\alpha_b=0.5$), 9:PIXELIZED_D4_A ($\alpha_b=0.5$), 10:PIXELIZED_D8_A ($\alpha_b=0.5$), 11:PIXELIZED_A ($ismax=True$, $D_a=Z_b$), 12:BLURRED_A ($\alpha_b=0.5$), and 13:BLURRED_A ($ismax=True$, $K_a=Z_b$).
  • Figure 4: Impact of the $\alpha_r$ on images with different resolutions from the VISPR dataset. (Top to bottom) image resolution $Z$: [$160 \times 120$], [$320 \times 240$], and [$1280 \times 960$]. (Left to right) 1:RAW_IMAGE, 2:PIXELIZED_D4_A ($\alpha_b=0.5$, $\alpha_r=0.5$), 3:PIXELIZED_D4_A ($\alpha_b=0.5$, $\alpha_r=Z/Z_{\text{ref}}$), 4:PIXELIZED_A ($ismax=True$, $D_a=Z_b$), 5:BLURRED_A ($\alpha_b=0.5$, $\alpha_r=0.5$), 6:BLURRED_A ($\alpha_b=0.5$, $\alpha_r=Z/Z_{\text{ref}}$), 7:BLURRED_A ($\alpha_b=0.5$, $\alpha_r=Z/Z_{\text{ref}}$, $isfullblur=True$), and 8:BLURRED_A ($ismax=True$, $K_a=Z_b$). The reference image size is set to $Z_{\text{ref}}=[320 \times 240]$.
  • Figure 5: Images consisting of persons at different depths on sample images from the VISPR dataset: (left to right) RAW_IMAGE, PIXELIZED_D4, PIXELIZED_D4_A, BLURRED and BLURRED_A.
  • ...and 3 more figures