Nav-YOLO: A Lightweight and Efficient Object Detection Model for Real-Time Indoor Navigation on Mobile Platforms

Cheng Su, Litao Zhu, Wen Dai, Jin Zhou, Jialiang Wang, Yucheng Mao, Jiangbing SunISPRS Int. J. Geo Inf 2025 · DOI

TL;DR

This paper introduces Nav-YOLO, a novel lightweight object detection model built upon YOLOv8n, specifically designed for real-time indoor navigation on mobile devices. By optimizing the backbone, neck, and detection head, Nav-YOLO achieves a superior balance of accuracy, speed, and computational efficiency, making it a practical solution for visually impaired navigation and other mobile applications.

What Problem Does This Paper Address?

Deploying deep learning-based object detection models on mobile platforms for tasks like indoor navigation presents a significant challenge. These models are often too large, computationally intensive, and slow for real-time performance on resource-constrained devices. This limits their practical application, especially for assistive technologies for visually impaired individuals who rely on accurate and immediate environmental understanding. The paper aims to bridge this gap by developing a highly efficient and lightweight object detection model that can run effectively on mobile devices.

Methods

The core of Nav-YOLO is a modified YOLOv8n architecture. Key innovations include:

  • Lightweight Backbone: Replacing YOLOv8n's backbone with ShuffleNetv2 significantly reduces model parameters and computational overhead while maintaining high accuracy. ShuffleNetv2 achieves this through depthwise separable convolutions and channel shuffling.
  • Slim-Neck Architecture: The neck component is redesigned using GSConv blocks and VoV-GSCSP modules. This "Slim-Neck" structure enhances feature fusion efficiency and adaptability for resource-constrained devices.
  • HyCTAS Detection Head: The detection head is upgraded using the Hybrid Convolutional Transformer Architecture Search (HyCTAS) technique. This module combines a memory-efficient self-attention mechanism with lightweight convolutional layers to optimize feature importance and reduce computational burden.
  • Inner-IoU Loss: To accelerate training convergence, the paper adopts the Inner-IoU loss function for bounding box regression. This method introduces a scale ratio parameter to manage auxiliary bounding boxes, leading to faster optimization.

Datasets

  • PASCAL VOC 2007 and VOC 2012: These standard datasets were used for training and evaluating the model's performance. They are widely used for object detection and semantic segmentation tasks.

Key Insights and Ideas

  • Systematic Co-design is Crucial: The paper emphasizes that simply combining existing lightweight components (like ShuffleNetv2, Slim-Neck, HyCTAS) isn't enough. Nav-YOLO's success comes from systematically co-designing these modules to complement each other, achieving an optimal trade-off between accuracy, speed, and computational cost.
  • Precision-Recall Trade-off in Lightweight Models: Integrating the Slim-Neck module led to a notable increase in precision but a slight decrease in recall. This is attributed to the module's bias towards high-confidence detections, prioritizing precision at the expense of recall, a common characteristic of feature-compressed architectures.
  • COS Metric for Holistic Evaluation: The introduction of the Composite Optimization Score (COS) provides a comprehensive evaluation framework that balances detection performance (Precision, Recall, mAP) with computational efficiency (model size, FLOPs), making it ideal for assessing mobile-friendly models.
  • Hardware Fragmentation is a Challenge: The deployment section highlights significant challenges due to hardware fragmentation across different mobile SoCs, impacting consistent performance and underscoring the need for hardware-aware optimization.

Key Results

  • Improved Accuracy: Nav-YOLO achieved significant improvements over the baseline YOLOv8n on the PASCAL VOC 2007 dataset, with mAP50 increasing by 10.3%. On VOC2012, it showed a 5% advantage in mAP50 over YOLOv8n.
  • Efficiency Gains: The model achieved substantial reductions in computational load. For instance, it reduced FLOPs by 15.9% compared to YOLOv8n.
  • Competitive Model Size: Nav-YOLO maintained a compact model size, comparable to or even smaller than YOLOv8n in some comparisons.
  • Real-Time Mobile Performance: Experimental results on an Android application demonstrated that Nav-YOLO achieves real-time detection speeds comparable to YOLOv8n, with some scenarios even showing faster inference, while maintaining superior detection accuracy.

Limitations and Future Work

  • Dataset Limitations: The authors acknowledge that PASCAL VOC datasets have limited category diversity and biases towards common scenes. This may lead to diminished generalization capabilities when applied to more diverse indoor environments with challenging conditions like dynamic lighting, occlusions, and dense clutter.
  • Misclassification of Similar Objects: The model occasionally misclassifies visually similar objects due to dataset constraints and architectural limitations that emphasize local features over global context.
  • Future Work:
    • Evaluate Nav-YOLO on more comprehensive, navigation-specific datasets that include varied illumination, obstructions, and a wider range of object categories.
  • Improve the model's ability to discriminate semantically similar classes through enhanced data curation and architectural refinement.
  • Expand training datasets to include more challenging navigation scenarios (complex urban obstacles, indoor environments) to enhance generalization.

Summary updated

BibTeX
@article{su2025nav,
  author = {Cheng Su and Litao Zhu and Wen Dai and Jin Zhou and Jialiang Wang and Yucheng Mao and Jiangbing Sun},
  title = {{Nav-YOLO: A Lightweight and Efficient Object Detection Model for Real-Time Indoor Navigation on Mobile Platforms}},
  journal = {ISPRS Int. J. Geo Inf.},
  year = {2025},
  doi = {10.3390/ijgi14090364},
}
Figure