High-Accuracy Model
The model achieved a 95% accuracy rate in pallet counting, meeting the client's requirements.
Technologies used
PyTorch
NVIDIA CUDA
OpenCV
NVIDIA Tensorrt
NVIDIA Deepstream
CUDNN
DeepSort
Gstreamer
ONNX Runtime
Ultralytics YOLO
Motivation
The client spent significant time manually counting 200–300 pallets daily, which was slow and prone to errors. The project aimed to automate this process with a computer vision system running offline on an edge device, capable of handling cluttered pallets and multiple camera feeds in real-time. This solution reduced human errors, saved time, and ensured reliable inventory tracking with scalability for future warehouse expansion.
Main challenges
Challenge 01
Offline Processing on Edge Device
The warehouse had limited or no internet connection, which meant the system had to run entirely on a local device. The client needed fast, reliable pallet counting without relying on the cloud. We optimized a deep learning model to run on an NVIDIA Jetson, using CUDA, cuDNN, and TensorRT for hardware acceleration, so the system could process video streams and count pallets in real-time right on the edge device.
#1
Challenge 02
Cluttered Pallet Arrangements
Pallets in the warehouse were often stacked irregularly or overlapping, making them hard to detect and count accurately. Manual counting was slow and error-prone. To solve this, we used YOLOv8 for object detection, OpenCV for preprocessing, and DeepSORT for tracking, enabling the system to reliably recognize and track pallets even in messy and complex arrangements.
#2
Our approach
Pre-Sale Analysis
We started by carefully studying the client's warehouse operations and the challenges of manual pallet counting. By observing current workflows and assessing the scale of the task, we proposed a solution based on computer vision and machine learning that could automate counting while ensuring high accuracy and offline operation.
Agile Development
The project followed an Agile methodology, allowing our team to develop iteratively, get feedback quickly, and continuously improve the system. This flexible approach ensured that we could adapt the model and system design as we learned more about real-world conditions in the warehouse.
Model Training
We collected 300–400 images daily from multiple cameras in the warehouse to train the deep learning model. By using real data from the client's environment, we ensured that the model could accurately detect and count pallets even under different lighting conditions, camera angles, and cluttered arrangements.
Edge Device Setup
We configured the NVIDIA Jetson edge device to process video streams locally, using GStreamer for efficient decoding and stream handling. This setup allowed the system to run completely offline, meeting the client's requirement for autonomous operation without any internet connection.
Model Optimization and Deployment
Once training was complete, we optimized the model using ONNX Runtime for faster inference on the edge device. The system was then deployed to count pallets in real-time, providing accurate, consistent results while freeing warehouse staff from manual counting and ensuring scalability for future needs.
Facing the same challenge?
Bring your complexity. We'll bring the plan. Select a convenient slot to start a conversation with our experts.
Schedule a callSolution
01
Offline Vision-Based Counting Module
Performs fully autonomous pallet counting without any internet connection, running entirely on an edge device. Handles video streams from multiple cameras and processes images in real-time to detect and count pallets accurately.
Key capabilities:
- Offline real-time processing on NVIDIA Jetson
- Supports multiple camera inputs simultaneously
- Handles varying lighting and cluttered pallet arrangements
02
Object Detection and Tracking Module
Detects pallets in complex and cluttered warehouse environments using YOLOv8, and tracks them reliably with DeepSORT to ensure accurate counts even when pallets overlap or move.
Key capabilities:
- High-accuracy object detection for pallets
- Real-time tracking of pallets across frames
- Handles irregular and overlapping pallet arrangements
03
Model Optimization and Deployment Module
Optimizes the trained model for fast inference on edge devices using ONNX Runtime and TensorRT, ensuring efficient GPU utilization and consistent real-time performance.
Key capabilities:
- ONNX Runtime and TensorRT optimization for speed
- Hardware-accelerated GPU inference using CUDA/cuDNN
- Seamless deployment to NVIDIA Jetson devices
04
Scalable System Architecture
Provides a flexible system design that allows for adding more cameras, expanding warehouse areas, or updating the detection model without affecting existing operations.
Key capabilities:
- Modular architecture for easy scaling
- Future-proof design for model updates
- Supports expansion to multiple warehouse zones
Results & business impact
Offline Processing
The solution runs entirely offline on the NVIDIA Jetson device, eliminating the need for an internet connection.
Increased Efficiency
Automation saved significant time for warehouse employees, reduced human error, and enhanced overall operational efficiency.
Scalability
The solution is easily adaptable and scalable, with the ability to update the model as needed for future improvements.
