RMP-YOLO: Robust Multi-Scale Pedestrian Detection for Dense Scenarios

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Autonomous driving systems require advanced pedestrian detection.
• Current dense pedestrian detection methods face performance limitations, especially with small, occluded, or low-light objects.

Purpose of the Study:

• To develop a robust and lightweight pedestrian detection algorithm (RMP-YOLO) for autonomous driving.
• To improve the detection of challenging pedestrian scenarios like small, occluded, and low-light conditions.

Main Methods:

• Utilized RFAConv in the backbone network, combining standard convolution, attention mechanisms, and group convolution for feature extraction.
• Integrated MobileViTv3 to merge Convolutional Neural Networks (CNNs) with Transformers, enhancing feature fusion and local representation.
• Employed the PIoUv2 loss function for precise bounding-box regression, particularly for small pedestrians.

Main Results:

• RMP-YOLO achieved a 1.3% mAP@0.5 improvement on a custom dataset and 0.91% on the WiderPerson dataset.
• The algorithm demonstrates high efficiency with only 3.71 million parameters and 6.29 GFLOPs.
• Significantly reduced detection errors for small-scale pedestrians in crowded environments.

Conclusions:

• RMP-YOLO offers a computationally efficient and accurate solution for dense pedestrian detection in autonomous driving.
• The proposed method effectively addresses performance bottlenecks in detecting challenging pedestrian targets.
• Meets deployment requirements for systems with limited computational power and high precision demands.