Abstract
This paper presents a real-time, CPU-efficient pedestrian detection system designed for human-aware robot navigation. By cascading an Aggregate Channel Features (ACF) detector with a fine-tuned Convolutional Neural Network (CNN), the system achieves robust person tracking on standard hardware without requiring expensive GPUs. The detected pedestrian positions are integrated into a navigation module using asymmetric Gaussian cost functions to model social norms such as personal space and overtaking rules. Experimental results demonstrate a log-average miss rate of 15.13% and real-time performance at 10 FPS on a standard CPU.
Introduction
For mobile robots to move naturally in human environments, they must do more than just avoid collisions; they must respect social norms, such as maintaining personal space (proxemics) and following "rules of the road" like overtaking on the left. The authors identify two primary challenges: the high computational cost of state-of-the-art deep learning detectors and the need for a standardized way to translate pedestrian positions into "cost functions" for socially acceptable path planning.
Methodology
The Detection Cascade
Instead of running a heavy CNN across every possible location in an image, the authors propose a two-stage cascade approach.
- Stage 1 (ACF): The Aggregate Channel Features detector quickly identifies "promising" image regions. It is fast but prone to false positives.
- Stage 2 (CNN): A VGG-16 model, downscaled and fine-tuned for pedestrian detection, classifies only the regions proposed by the ACF.
This drastically reduces the number of windows the CNN must process, enabling real-time performance on a standard CPU. The CNN effectively filters out false positives while retaining true pedestrian detections.
Human-Aware Navigation (HAN)
Once pedestrians are detected and tracked (using a Kalman Filter and NNJPDA), their positions are projected onto a 2D world map. Social constraints are modeled as asymmetric Gaussian cost functions, which account for personal space (larger in front of a person), visibility (preventing passing too close behind seated individuals), and social interactions (e.g., space between a person and a TV). This asymmetry reflects human psychology; humans are more sensitive to intrusions in their field of vision than from behind.
Key Results
The ACF+CNN cascade achieved a log-average miss rate of 15.13% on the INRIA dataset, an improvement over the 16.83% achieved by ACF alone. The system reached approximately 10 FPS on a standard i7 CPU during real-world testing. In experiments with the "MBOT" platform, the robot successfully performed "slalom" maneuvers to avoid personal spaces and replanned its path when social constraints changed, such as when a person started watching TV.
Conclusion
This research demonstrates that selective search strategies can make deep learning feasible for real-time applications on CPU-based mobile robots. By combining efficient detection with asymmetric social modeling, robots can navigate human environments in a more natural and socially acceptable manner. Future work will explore learning social parameters from human behavior data and handling more complex crowd interactions.
Summary updated