Research on Visual Recognition of Service Robots Based on YOLO Algorithm

Bo Yuan, Jiajun Su, Changlong Dai, Jingjing Zha, Jialin Sun, Xinyi HouJournal of Computing and Electronic Information Management 2024 · DOI

Abstract

In modern society, based on the aging of population and the rapid development of intelligent technology, the demand for intelligent service robots is increasing, and visual recognition is one of the key technologies of service robots. The rapid development of deep learning technology provides new solutions for visual recognition. Existing target recognition algorithms have the problem of complex model and slow inference speed. According to the above problems, this paper proposes a visual recognition algorithm based on YOLO series, which is used for the visual recognition design scheme of service robot. It can ensure the detection accuracy, reduce the speed of the model complexity, speed up the reasoning, and determine whether to meet the actual needs. Through randomly selected pictures, the results show that the design scheme can effectively improve the detection accuracy and real-time of the service robot on targets, and enhance its autonomous service ability in various complex environments.

Introduction

As the global population ages, the demand for intelligent service robots—ranging from medical transporters to domestic assistants—is increasing significantly. For these robots to be effective, they must perceive and understand their surroundings in real-time. However, developers face a classic engineering trade-off: high-accuracy deep learning models are often computationally intensive and slow to run, while lightweight models may sacrifice reliability. This research explores the implementation of the YOLOv5s algorithm to provide service robots with efficient, real-time visual recognition, balancing detection accuracy with the low latency required for autonomous navigation.

Methodology

The study focuses on the YOLOv5 (version 5.0) framework, a single-stage object detection algorithm optimized for speed. The architecture is structured into four primary components: the Input, the Backbone (for feature extraction), the Neck (for multi-scale feature fusion), and the Head (for final prediction).

The architecture of the YOLOv5 model, illustrating the Backbone, Neck, and Head components.

Model Selection

The YOLOv5 series offers several variants: YOLOv5s, 5m, 5l, and 5x. While larger models provide higher accuracy, they require more computational resources and have higher inference latency. For service robots, real-time performance is paramount.

Comparison of YOLOv5 variants showing the trade-off between COCO AP value and GPU inference speed.

As shown in the performance comparison, the researchers selected YOLOv5s (Small). Despite having the lowest theoretical accuracy among the variants, its minimal parameter count and high inference speed make it the most viable candidate for real-time robotics applications where safety and fluid interaction depend on low latency.

Experimental Setup

The model was trained using the GIOU (Generalized Intersection over Union) loss function over 300 epochs. The development environment utilized PyTorch 2.3.0 and CUDA 12.3 on a system equipped with an Intel i7-9750H CPU and an NVIDIA RTX 3080 Ti GPU.

A custom dataset of 1,645 images was curated across four categories: Airplane (321), Bicycle (373), Bird (465), and Boat (486). Images were manually annotated using LabelImg, and the data was split into training and testing sets using an 8:2 ratio.

Results and Discussion

The experimental results demonstrated high performance across the test categories. The model achieved the following accuracies:

  • Bicycle: 95%
  • Bird: 92%
  • Ship/Boat: 87%
  • Aircraft/Airplane: 85%

The high accuracy for bicycles suggests that objects with distinct, high-contrast geometric features (like wheels) are easier for lightweight models to resolve. Conversely, aircraft silhouettes, which often have lower contrast against the sky, presented a greater challenge. This highlights that in robotics, "good enough" accuracy at high frame rates is often preferable to high accuracy at low frame rates.

Conclusion

The YOLOv5s design effectively improves the real-time detection capabilities of service robots, allowing them to maintain autonomous service in complex environments. However, challenges remain regarding the detection of very small or occluded targets. Future work will focus on enhancing environmental adaptability—such as performance in low-light or crowded settings—and tailoring models for specific robotic applications like medical or agricultural assistance.

Summary updated

Abstract

In modern society, based on the aging of population and the rapid development of intelligent technology, the demand for intelligent service robots is increasing, and visual recognition is one of the key technologies of service robots. The rapid development of deep learning technology provides new solutions for visual recognition. Existing target recognition algorithms have the problem of complex model and slow inference speed. According to the above problems, this paper proposes a visual recognition algorithm based on YOLO series, which is used for the visual recognition design scheme of service robot. It can ensure the detection accuracy, reduce the speed of the model complexity, speed up the reasoning, and determine whether to meet the actual needs. Through randomly selected pictures, the results show that the design scheme can effectively improve the detection accuracy and real-time of the service robot on targets, and enhance its autonomous service ability in various complex environments.
BibTeX
@article{yuan2024research,
  author = {Bo Yuan and Jiajun Su and Changlong Dai and Jingjing Zha and Jialin Sun and Xinyi Hou},
  title = {{Research on Visual Recognition of Service Robots Based on YOLO Algorithm}},
  journal = {Journal of Computing and Electronic Information Management},
  year = {2024},
  doi = {10.54097/w2qqng47},
}
Figure