Human Detection and Tracking for Video Surveillance: A Cognitive Science Approach

Vandit Gajjar, Yash Khandhediya, Ayesha GurnaniWorkshops @ ICCV 2017 · DOI · arXiv

Abstract

Traditional human detection in video surveillance often suffers from high computational costs and excessive false positives due to exhaustive search methods. This paper presents a novel approach that integrates visual saliency—a cognitive science principle modeling human attention—to optimize the detection pipeline. By utilizing a Deep Multi-Level Network (ML-Net) to generate saliency maps, the proposed method prunes the search space for Histogram of Oriented Gradients (HOG) and Support Vector Machine (SVM) classifiers. Experimental results on the OSU Color-Thermal Pedestrian Dataset demonstrate a significant increase in precision from 12.36% to 83.11% and a reduction in processing time from nearly 30 minutes to 23 seconds per frame.

Introduction

Automated video surveillance systems face a persistent challenge: the "needle in a haystack" problem. Standard detection algorithms typically rely on a "sliding window" approach, meticulously scanning every possible region of a high-resolution image to identify human patterns. In complex environments, this exhaustive search is computationally prohibitive and prone to false positives triggered by background clutter like shadows or architectural elements. This research seeks to bridge computer vision and cognitive science by teaching algorithms to prioritize regions that would naturally attract a human observer's attention.

Methodology

The proposed system employs a multi-stage pipeline designed to enhance efficiency and accuracy through background suppression:

  1. Saliency Mapping: A Deep Multi-Level Network (ML-Net) acts as a cognitive shortcut, generating a heat map where brighter pixels indicate areas of high visual interest.
  2. Salience-Windowing: The original RGB image is multiplied by the saliency map. This process effectively "dims" irrelevant background data while "highlighting" potential regions of interest (ROIs).
  3. Detection: A standard HOG feature extractor and SVM classifier are applied only to these highlighted regions. By suppressing the background, the classifier's exposure to noise is drastically reduced.
  4. Tracking: For post-analysis, the system utilizes k-means clustering on HOG feature vectors across consecutive frames to reconstruct individual motion paths.
The proposed detection pipeline integrates a saliency-windowing stage to focus the sliding window search.

Key Insights and Results

The core innovation lies in treating visual saliency as a sophisticated region proposal mechanism. By focusing computational resources only on salient regions, the system achieves a massive 76x speedup.

A salience-windowed image demonstrates how background suppression isolates potential human targets.

As shown in the table below, the transition from "Normal" exhaustive search to "Salience-windowed" detection yields dramatic improvements in precision and execution speed, albeit with a trade-off in recall.

Metric Normal Images Salience-windowed
Execution Time ~1768.29 s ~23.00 s
Precision 12.36% 83.11%
Recall 93.12% 41.27%

The sharp increase in precision (from 12% to 83%) indicates that the system is far more confident in its detections, effectively eliminating the "ghost" detections that plague traditional HOG/SVM implementations. However, the drop in recall to 41% suggests that the current saliency model may overlook individuals in low-contrast areas or near the frame boundaries.

Conclusion and Future Work

Integrating cognitive saliency into the detection pipeline offers a viable path toward real-time automated surveillance. While the current iteration significantly boosts precision and speed, future work will focus on improving recall. The authors plan to incorporate optical flow to capture motion dynamics and refine the ML-Net architecture to better handle edge-case saliency near image borders. Transitioning the tracking component from post-processing to an online, real-time system remains a primary objective for active security applications.

Summary updated

BibTeX
@inproceedings{gajjar2017human,
  author = {Vandit Gajjar and Yash Khandhediya and Ayesha Gurnani},
  title = {{Human Detection and Tracking for Video Surveillance: A Cognitive Science Approach}},
  booktitle = {Workshops @ ICCV},
  year = {2017},
  doi = {10.1109/iccvw.2017.330},
}
Figure