Source-linked AI summary
Real-Time Apple Detection System Using Embedded Systems With Hardware Accelerators: An Edge AI Application
Vittorio Mazzia, Francesco Salvetti, Aleem Khaliq, Marcello Chiaberge
TL;DR
Real-time orchard apple detection is constrained by environmental occlusion and the need for lightweight, fast embedded systems. The study adapts YOLOv3-tiny for small and occluded apples and evaluates it across edge-AI platforms, achieving improved detection performance and up to 30 fps, while leaving image registration unaddressed.
Problem
Apple detection in natural orchard environments is difficult because of illumination, shadows, branches, and leaf coverings, while real-time harvesting systems require direct, fast, and accurate recognition.
Method
The study modifies YOLOv3-tiny for small and heavily occluded objects, trains it on custom real-orchard data, and tests it on embedded edge-AI platforms using object-detection evaluation metrics.
Results
Up to 30 fps was achieved on the NVIDIA Jetson AGX Xavier in 30W mode, while the adapted architecture increased test-set mAP by up to 6.6%.
Takeaways & Limitations
The tested system shows prospects for real-time apple position and count detection with minimal power consumption across embedded platforms.
Takeaways & Limitations
Image registration was not directly addressed, and further work is focused on yield estimation using reliable apple counting.
Abstract
from arXiv · showhide
Real-time apple detection in orchards is one of the most effective ways of estimating apple yields, which helps in managing apple supplies more effectively. Traditional detection methods used highly computational machine learning algorithms with intensive hardware set up, which are not suitable for infield real-time apple detection due to their weight and power constraints. In this study, a real-time embedded solution inspired from "Edge AI" is proposed for apple detection with the implementation of YOLOv3-tiny algorithm on various embedded platforms such as Raspberry Pi 3 B+ in combination with Intel Movidius Neural Computing Stick (NCS), Nvidia's Jetson Nano and Jetson AGX Xavier. Data set for training were compiled using acquired images during field survey of apple orchard situated in the north region of Italy, and images used for testing were taken from widely used google data set by filtering out the images containing apples in different scenes to ensure the robustness of the algorithm. The proposed study adapts YOLOv3-tiny architecture to detect small objects. It shows the feasibility of deployment of the customized model on cheap and power-efficient embedded hardware without compromising mean average detection accuracy (83.64%) and achieved frame rate up to 30 fps even for the difficult scenarios such as overlapping apples, complex background, less exposure of apple due to leaves and branches. Furthermore, the proposed embedded solution can be deployed on the unmanned ground vehicles to detect, count, and measure the size of the apples in real-time to help the farmers and agronomists in their decision making and management skills.
I. INTRODUCTION
Real-time apple detection is important for crop-load and yield assessment, but orchard conditions and embedded-device constraints make accurate, efficient detection difficult.
- Fruit counting and size estimation support crop-load assessment and yield mapping in tree fruit crops.
- 87?
- Variable lighting, occlusion, branches, and leaves make robust apple detection challenging in natural environments.
- Faster R-CNN achieves strong fruit-detection accuracy but uses region proposal and classification networks with substantial computational complexity.
- Low-power embedded platforms raise concerns about CNN efficiency, motivating model compression techniques such as pruning and quantization.
A. AI ON THE EDGE AND RELATED WORK
Edge AI performs computation locally to reduce cloud-related constraints, while embedded accelerators provide hardware options for real-time agricultural detection.
- Edge AI performs computations locally on data acquired from devices or sensors.
- Edge-AI systems must balance high output accuracy with low power consumption.
- NVIDIA Jetson platforms combine CPUs with CUDA-capable GPUs to accelerate machine-learning tasks in lightweight, low-power hardware.
- The study deploys modified YOLOv3-tiny on Raspberry Pi 3 B+ with Intel NCS, Jetson Nano, and Jetson AGX Xavier for real-time apple detection.
II. MATERIALS AND DATA
The study combines orchard imagery with transfer learning to retrain and fine-tune YOLOv3-tiny for small-object apple detection on embedded devices.
- An orchard-derived custom dataset is used for training, followed by transfer learning and fine-tuning of a customized YOLOv3-tiny network.
- The resulting network is benchmarked on Apple-class images from the OIDv4 dataset across training, validation, and testing data.
A. DATA SET DESCRIPTION
The dataset covers Braeburn and Fuji apples photographed in northern Italian orchards across fruit-overlap and lighting conditions.
- The study considers Braeburn and Fuji apples, described as common varieties in northern Italy.
- Images were acquired from randomly selected healthy apple trees using an 18-megapixel reflex camera during different times and days in September.
- The acquisition includes separate, overlapping, and occluded fruits under variable lighting conditions.
B. HARDWARE DESCRIPTION
The study separates high-compute model training from real-time inference on embedded hardware, evaluating Raspberry Pi–NCS combinations and NVIDIA Jetson platforms for edge deployment.
- Edge AI deployment: Edge AI performs inference locally on embedded systems while training remains on a dedicated workstation.The model weights obtained during workstation training are deployed to target hardware for execution.
- Analyzed platforms: The evaluated platforms include Raspberry Pi 3 B+ with Intel Movidius accelerators, NVIDIA Jetson Nano, and NVIDIA Jetson AGX Xavier.The study examines both USB accelerator configurations and embedded GPU platforms.
- Analyzed platforms: Intel Neural Compute Sticks use Myriad VPU processors to accelerate neural-network inference through a USB 3.0 interface.Both first- and second-generation sticks are tested with a Raspberry Pi 3 B+ host.
- Analyzed platforms: NVIDIA Jetson boards combine an embedded computer with a dedicated GPU for hardware acceleration.The investigated boards are the AGX Xavier and Jetson Nano.
- Platform characteristics: The AGX Xavier provides substantially greater accelerator capability than the Nano, while the Nano targets reduced size, power consumption, and price.The AGX Xavier reaches about 16 TFLOPS peak performance, whereas the Nano reaches 472 GFLOPs and operates at 5 W or 10 W.
III. METHODOLOGY AND ARCHITECTURE FRAMEWORK
The methodology uses YOLO as a fast, compact fully convolutional detector and selects its lighter tiny variant for embedded real-time apple detection, with modifications aimed at small objects.
- YOLO framework: YOLO takes a raw image as input and outputs bounding boxes with the classes of recognized objects.Its fully convolutional design supports real-time object detection.
- YOLOv3-tiny: YOLOv3-tiny simplifies and optimizes the YOLO structure while retaining much of its accuracy and inference speed.These characteristics make the tiny variant suitable for edge AI applications on embedded systems.
- Small-object adaptation: The study modifies YOLOv3-tiny to improve detection of smaller objects such as apples in orchard scenes.The architecture is adapted specifically for small-object detection on embedded devices.
A. ARCHITETURE OF THE ORIGINAL FRAMEWORK
The original YOLOv3-tiny framework is a fully convolutional network built around the darknet-19 feature extractor; the study adds multiscale prediction to improve small-object detection.
- Original framework: YOLOv3-tiny uses convolutional layers throughout and can accept inputs of different sizes during and after training.Its backbone, darknet-19, hierarchically extracts features from raw input pixels.
- Architecture modification: The modified architecture performs predictions at three different scales by adding a detection scale with stride 8.The added scale is intended to improve recall and robustness to scale variations.
- Evaluation setup: The architecture comparison evaluates precision and recall using confidence threshold c = 0.25 and IOUtarget = 0.5.These settings are used when comparing the original and customized models on excluded images.
- Training setup: Training uses an NVIDIA RTX 2080 Ti workstation with 64GB of DDR4 SDRAM and the TensorFlow framework with CUDA 10.The reported average training time is one hour.
B. QUANTITATIVE RESULTS: MODEL PERFORMANCE
The customized YOLOv3-tiny architecture improves apple detection for small fruits and is evaluated across embedded platforms using detection accuracy, power, and frame-rate measures.
- Model performance: 83.64% mean average detection accuracy is reported for the proposed apple detector.
- Model performance: Up to 6.6% mAP improvement is achieved over the retrained original YOLOv3-tiny architecture.The comparison uses the same apple-detection training methodology and test dataset.
- Embedded implementation: The evaluation measures frame rate and absorbed power after steady-state execution across Jetson AGX Xavier, Jetson Nano, and Raspberry Pi 3B+ platforms.Jetson boards are tested across their available working conditions.
- Embedded implementation: 30 fps is reached by the Jetson AGX Xavier in 30W mode, while the Jetson Nano reaches 8 fps in 10W mode.The Raspberry Pi and Intel USB accelerators perform more slowly; USB accelerators remain below 5 fps in the best case.
- Robustness: Qualitative tests show detection across varying illumination, viewpoint, scale, occlusion, and background clutter, including small apples in difficult lighting.These conditions are evaluated on additional images from the study site.
- Platform comparison: The Jetson Nano offers the best balance between performance and cost, whereas the AGX Xavier is unsuitable for low-cost solutions despite its higher price/fps ratio.The Intel Neural Sticks also require an additional embedded computer, increasing final cost.
D. QUALITATIVE RESULTS AND COMPARISON
The adapted network detects apples across scales and difficult illumination conditions, with especially strong gains in recall for very small fruits compared with the original architecture.
- The network detects apples at different scales, including very small apples in poor lighting conditions.Test images excluded from training covered varied illumination and contrast.
- The adapted architecture strongly increases recall for very small fruits compared with the original architecture.The comparison used precision and recall with confidence threshold c = 0.25 and IOUtarget = 0.5.
V. CONCLUSION
The study develops and evaluates a real-time apple detector for edge AI devices by adapting YOLOv3-tiny for small and occluded objects. The system improves recall and precision across disparate target sizes and supports real-time detection with low power consumption.
- The modified YOLOv3-tiny system was developed and tested on several edge AI devices for small and largely occluded apples.The model was trained on a custom orchard dataset.
- The system improves recall and precision when apple targets have disparate sizes.
- The experiments evaluated inference speed and power consumption across the selected embedded solutions.The results indicate prospects for real-time apple positions and detection counts with minimal power consumption.