This repository provides the implementation and dataset for the paper "PPE-Det: Lightweight Object Detection for Edge Devices", submitted to the 2nd International Conference on Intelligent Systems, Blockchain, and Communication Technologies (ISBCom) 2025.
The project focuses on evaluating lightweight object detection models for detecting helmets and coveralls (red, blue, yellow) in industrial environments. The models are optimized for Raspberry Pi 5 using PyTorch and NCNN frameworks.
The PPE-Det dataset comprises approximately 5,200 images collected to facilitate Personal Protective Equipment (PPE) detection in industrial environments. The primary objective of this PPE-Det dataset is to enhance the accuracy of industrial models by providing diverse object detection samples in real-world scenarios. The images were extracted from a recorded video in an Egyptian industrial setting and split into frames at a rate of 2 frames per second (fps).
The PPE-Det dataset utilizes bounding box annotations to enable object detection. The annotations follow the YOLO format, where each labeled instance is represented by a class ID and four normalized values (center x, center y, width, height) defining the bounding box.
All images were initially auto-labeled using an object detection model. The auto-generated labels were then manually reviewed and corrected for accuracy. All images were resized so that the largest dimension is 640 pixels, while preserving the aspect ratio.
The PPE-Det dataset contains object instances categorized across three distance ranges: Far, Mid, and Near, representing different levels of detail and resolution. The three distance ranges are quantified as follows:
- Near Range: Approximately 2.5-3m
- Mid Range: 4-5m
- Far Range: 8-10m
These distances refer to the camera height from the ground.
| Class ID | Description | Far-Range | Mid-Range | Near-Range | Total Instances |
|---|---|---|---|---|---|
| 0 | Helmet | 1,884 | 3,593 | 1,878 | 7,355 |
| 1 | Blue Coverall | 1,412 | 553 | 385 | 2,350 |
| 2 | Red Coverall | 338 | 1,604 | 855 | 2,797 |
| 3 | Yellow Coverall | 341 | 1,669 | 905 | 2,915 |
| Total | - | 3,975 | 7,419 | 4,023 | 15,417 |
| Distance Range | Number of Images |
|---|---|
| Far-Range | 1,695 |
| Mid-Range | 1,676 |
| Near-Range | 1,840 |
| Total | 5,211 |
The data is split into three subsets: 70% for training, 15% for validation, and 15% for testing.
The dataset is hierarchically organized. At the top level, the dataset is divided into three main splits: train, val, and test. Each of these splits is categorized into three subdirectories—far, mid, and near—based on object distance. Each of these subdirectories contains two subfolders: images/ (JPG format) and labels/ (TXT in YOLO format).
dataset/
│── train/
│ ├── far/
│ │ ├── images/ (JPG)
│ │ ├── labels/ (TXT in YOLO format)
│ ├── mid/
│ ├── near/
Each YOLO annotation file follows a structured format where each line represents an object in the image:
<class_id> <center_x> <center_y> <width> <height>
Class ID Mapping:
- 0: Helmet
- 1: Blue Coverall
- 2: Red Coverall
- 3: Yellow Coverall
The PPE-Det dataset can be accessed via this Google Drive Link.
Below are sample images representing the three distance ranges:
- Far Range:
- Mid Range:
- Near Range:
The following models were trained and evaluated on the PPE-Det dataset for object detection:
| Model | Architecture | Description |
|---|---|---|
| YOLOv5n | Ultra-lightweight | Designed for fast inference on edge devices. |
| YOLOv8n | Next-gen YOLO | Improved accuracy and efficiency over previous YOLO versions. |
| YOLOv9t | Transformer-enhanced YOLO | Incorporates transformer layers for better feature representation. |
| YOLOv10n | Enhanced lightweight | Optimized for better performance on low-power devices. |
| YOLOv11n | Further optimized | Improved speed-accuracy tradeoff compared to YOLOv10n. |
| YOLOX-Nano | Nano variant | Efficient object detection designed specifically for mobile and edge devices. |
| NanoDet-M | Lightweight anchor-free model | Specializes in fast and efficient object detection. |
These models were evaluated based on metrics such as Mean Average Precision (mAP) and Inference Time. Performance evaluations were conducted on a Raspberry Pi 5 to simulate real-world edge deployment scenarios.
| Model | Parameters (M) | GFLOPS |
|---|---|---|
| YOLOv5n | 2.5 | 7.1 |
| YOLOv8n | 3 | 8.1 |
| YOLOv9t | 1.97 | 7.6 |
| YOLOv10n | 2.7 | 8.2 |
| YOLOv11n | 2.58 | 6.3 |
| YOLOX-Nano | 0.9 | 1.08 |
| NanoDet-M | 0.95 | 0.72 |
| MobileNet-SSD | 2.25 | 0.85 |
-
YOLOv5n: Uses Cross-Stage Partial Networks (CSPNet) for enhanced gradient flow and reduced computational redundancy. Incorporates depthwise separable convolutions and a refined PANet-based neck for balanced detection speed and accuracy.
-
YOLOv8n: Implements an anchor-free detection paradigm, improving localization precision with adaptive spatial feature fusion for efficient multi-scale learning.
-
YOLOv9t: Integrates Transformer-based modules to enhance long-range dependencies and feature reuse. Utilizes depthwise separable convolutions to boost computational efficiency.
-
YOLOv10n: Combines convolutional and attention-based techniques with a streamlined detection head, improving detection robustness and reducing parameter complexity.
-
YOLOv11n: Incorporates enhanced multi-scale feature fusion and knowledge distillation mechanisms. Optimized activation functions and lightweight attention modules improve precision.
-
YOLOX-Nano: Employs a decoupled head structure with an anchor-free approach, simplifying detection and enhancing adaptability to varying object sizes.
-
NanoDet-M: Utilizes a one-stage anchor-free framework with an enhanced Feature Pyramid Network (FPN) for improved multi-scale aggregation and efficient localization.
-
MobileNet-SSD: Uses MobileNetV3 backbone with SSDLite detection head, incorporating Squeeze-and-Excitation (SE) modules, h-swish activation, and Neural Architecture Search (NAS) for better accuracy and latency balance.
The Raspberry Pi 5 was chosen for this experiment due to its balance of computational power and affordability, making it suitable for real-time object detection. Its specifications are:
- Processor: Quad-core ARM Cortex-A76 (2.4 GHz)
- Memory: 8GB LPDDR4X RAM
- Storage: 64GB MicroSD card
- Operating System: Ubuntu Server
Compared to higher-cost alternatives like the NVIDIA Jetson, the Raspberry Pi 5 provides a cost-effective, widely accessible solution with strong community support.
The experiments were conducted in a controlled environment to ensure consistency across evaluations. Object detection models were deployed using Docker containers, running exclusively on CPU-based inference. Two different Docker images were utilized:
pytorch/torchserve:0.12.0-cpufor non-standard YOLO models.ultralytics/ultralytics:latest-arm64for standard YOLO models.
The dataset was divided into training, validation, and test sets with a 70:15:15 ratio for balanced evaluation.
The models were trained independently using the Adam optimizer with a learning rate of
Detection accuracy was evaluated using mAP@50 (IoU threshold at 50%) and mAP@50-95 (average precision over thresholds from 50% to 95%). YOLO-based models achieved the highest accuracy, particularly YOLOv9t and YOLOv11n, with mAP@50 and mAP@50-95 scores of 0.973 and 0.78, respectively.
| Model | Near (mAP50 / mAP50-95) | Mid (mAP50 / mAP50-95) | Far (mAP50 / mAP50-95) | All (mAP50 / mAP50-95) |
|---|---|---|---|---|
| YOLOv5n | 0.986 / 0.853 | 0.957 / 0.724 | 0.924 / 0.655 | 0.956 / 0.734 |
| YOLOv8n | 0.989 / 0.858 | 0.953 / 0.724 | 0.928 / 0.666 | 0.96 / 0.738 |
| YOLOv9t | 0.99 / 0.892 | 0.964 / 0.769 | 0.951 / 0.726 | 0.973 / 0.778 |
| YOLOv10n | 0.983 / 0.843 | 0.955 / 0.71 | 0.941 / 0.688 | 0.963 / 0.73 |
| YOLOv11n | 0.992 / 0.894 | 0.966 / 0.771 | 0.96 / 0.72 | 0.973 / 0.78 |
| YOLOX-nano | 0.939 / 0.697 | 0.864 / 0.552 | 0.828 / 0.479 | 0.873 / 0.578 |
| Nanodet-M | 0.876 / 0.656 | 0.775 / 0.513 | 0.739 / 0.428 | 0.785 / 0.505 |
| MobileNet-SSD | 0.744 / 0.46 | 0.429 / 0.193 | 0.412 / 0.163 | 0.536 / 0.28 |
Inference latency was measured on a Raspberry Pi 5 using PyTorch and NCNN formats. The NCNN backend significantly reduced latency by over 60% for some models. The results are shown below:
| Model | PyTorch (Near) (ms) | PyTorch (Mid) (ms) | PyTorch (Far) (ms) | PyTorch (All) (ms) | NCNN (Near) (ms) | NCNN (Mid) (ms) | NCNN (Far) (ms) | NCNN (All) (ms) |
|---|---|---|---|---|---|---|---|---|
| YOLOv5n | 298.55 | 301.33 | 290.78 | 296.22 | 104.97 | 106.55 | 103.78 | 105.25 |
| YOLOv8n | 307.45 | 309.12 | 300.87 | 305.81 | 106.98 | 107.55 | 104.61 | 106.38 |
| YOLOv9t | 358.11 | 359.79 | 350.25 | 356.05 | 125.33 | 126.45 | 123.65 | 124.81 |
| YOLOv10n | 354.23 | 356.91 | 346.12 | 352.42 | - | - | - | - |
| YOLOv11n | 324.55 | 326.78 | 315.89 | 322.41 | 111.45 | 112.68 | 108.87 | 111.00 |
| YOLOX-nano | 240.33 | 242.10 | 235.82 | 238.75 | - | - | - | - |
| Nanodet-M | 149.55 | 150.78 | 144.21 | 148.18 | - | - | - | - |
| MobileNet-SSD | 236.45 | 237.98 | 230.22 | 234.63 | - | - | - | - |
- YOLOv9t and YOLOv11n outperform others in detection accuracy, particularly at far distances.
- Detection accuracy decreases with distance but remains strong for YOLOv9t and YOLOv11n.
- Smaller models (YOLOX-Nano, NanoDet-M, MobileNet-SSD) show lower accuracy due to reduced size and computational complexity.
- The NCNN backend provides a significant reduction in inference latency, especially for YOLOv5n and YOLOv8n.
- Inference latency is not significantly affected by camera distance.
To reproduce the results of this project, there are two available methods:
- YOLO Versions: Located inside the directory
/code/yolo_versions. - Non-YOLO Versions: Located inside the directory
/code/non_yolo_versions.
Please refer to the respective directories for detailed instructions and implementation.
This work is submitted to the ISBcom 2025 conference. The authors with equal contribution are:
- Abdelaziz Essam
- Ahmed Yasser
- Ahmed Hassan
The project was conducted under the supervision of Prof. Walid Gomaa at the Computer Science Engineering (CSE) Department, Egypt Japan University of Science and Technology (EJUST).