TL;DR
This paper presents a novel, real-time head detection method using Kinect depth data, focusing on the elliptical shape of heads. This approach significantly speeds up driving fatigue detection systems by replacing computationally expensive color-based face detection with a faster depth-based head localization step.
What Problem Does This Paper Address?
Accurate and fast head detection is crucial for many computer vision applications, including driving fatigue detection. Traditional methods often rely on color images, which are susceptible to varying lighting conditions. While depth cameras like the Kinect offer a more robust solution, existing depth-based head detection methods can be slow or struggle with noise and different head poses. This paper aims to bridge this gap by proposing a real-time, robust head detection method using Kinect depth data that is less sensitive to lighting and can achieve high speeds, making it suitable for real-time applications like driver monitoring.
Methods
The proposed method leverages the assumption that human heads have an approximately elliptical shape and continuous depth values. It comprises three key steps:
- Depth Image Split: The depth image is segmented into regions based on depth value changes. Pixels with depth differences below a certain threshold are grouped into connected components. Regions that are too large or too small to be a human head are discarded.
- Contour Extraction: For each segmented region, the algorithm extracts its contour points. This is achieved by a "walking" procedure that traverses the boundary of the region, identifying points and their outward normal directions.
- Ellipse Fitting: An ellipse fitting algorithm is applied to the extracted contour points of each region. The similarity between the region and the fitted ellipse is calculated using a normalized average offset. Regions with an elliptical shape and a fitness score below a certain threshold are identified as human heads.
The authors highlight that their method avoids computationally intensive template matching, which is a bottleneck in some prior works, contributing to its speed.
Datasets
The paper evaluates its head detection method on the RGB-D database provided by Hg et al. [6]. This database contains 31 individuals, each with 17 poses, captured with a Microsoft Kinect. For each pose, color and depth images were recorded three times. The authors specifically used the depth images for their tests, totaling 1581 depth images with a resolution of 640x480.
Key Insights and Ideas
- Depth Data for Robustness: Utilizing depth information from Kinect significantly improves robustness to varying lighting conditions, a common challenge in driving scenarios.
- Elliptical Shape as a Strong Cue: The assumption of an elliptical head shape, combined with continuous depth, proves to be a powerful and efficient feature for head detection.
- Speed Through Simplicity: By avoiding complex template matching and focusing on contour and ellipse fitting, the method achieves remarkable speed improvements over previous approaches.
- Integration for Practicality: The successful integration of this fast head detection into a driving fatigue system demonstrates its real-world applicability and the significant speedup it provides.
Key Results
The proposed head detection method achieved a satisfactory detection rate across most individuals in the tested database, showing robustness to head pose. The average detection rate was 83.05%. Notably, the method achieved a speed of approximately 170ms per depth image (640x480 resolution) on a Pentium T4400 2.2GHz CPU with 4GB RAM, which is about twice as fast as a comparable method [8]. The authors also reported that integrating this method into their driving fatigue detection system reduced the overall processing time for head and face detection from 1800ms to approximately 240ms (170ms for head detection + 70ms for face detection within the head region).
Limitations and Future Work
The authors acknowledge that their assumptions about elliptical head shape and continuous depth values do not hold for every individual, particularly those with very bushy hair or beards, which can lead to a significant drop in detection rate for such cases.
For future work, they plan to:
- Improve the method to better handle individuals with very bushy hair or beards.
- Add more features to the driving fatigue detection system to achieve more accurate results.
Summary updated