Improving pedestrian safety using combined HOG and Haar partial detection in mobile systems

Area of Science:

• Computer Vision
• Mobile Systems
• Automotive Safety

Background:

• Traditional Histogram of Oriented Gradients (HOG) detection, while effective, lacks the speed required for real-time mobile applications.
• Existing pedestrian detection methods often struggle to balance detection rate and processing speed on resource-constrained mobile platforms.

Purpose of the Study:

• To develop a novel algorithm for mobile systems capable of near-real-time pedestrian detection and collision warning.
• To enhance pedestrian detection speed and accuracy for automotive safety applications.

Main Methods:

• A hybrid approach combining partial Haar transform pre-detection with Histogram of Oriented Gradients (HOG) detection.
• Implementation of a score-based confidence level through multi-detection merging for prioritized outcomes.

Main Results:

• Achieved a significant increase in detection speed from 46 to 76 frames per second (fps).
• Improved the detection rate from 80% to 91% compared to standard HOG detection.
• Enhanced result confidence via merged detections and assigned scores, enabling tiered warning levels.

Conclusions:

• The proposed algorithm demonstrates superior performance in both detection rate and speed compared to state-of-the-art methods.
• The algorithm is well-suited for mobile systems, providing a viable solution for warning drivers of potential pedestrian collisions.