How we leveraged edge computing to improve warehouse management optimization
We joined forces with PPS-AI, a leading solutions service provider from Australia. Their business specializes in automation from identification and recognition in logistics and parcel industries.
We helped them introduce Computer Vision technology to optimize a retailer’s operations and warehouse management. We built three models combining different AI approaches, such as object detection, segmentation and tracking.
The client wanted a consistent and high-quality code to optimize warehouse management by tracking the movements inside of it, calculating the available and unavailable spaces and monitoring the movements of trucks in a designated area.
We created our own dataset by collecting and labeling images from the warehouse using various cameras. In some cases, this involved developing specialized labeling tools for efficient annotation or finding the best positioning for the cameras to capture optimal data. We manually labeled key elements such as pallets, merchandise, workers, forklifts, and trucks to create labeled datasets. Furthermore, we implemented careful boundary definition and post-processing steps to enhance model accuracy.
We customized our model selection for each challenge. For pallet counting, we started with object detection using bounding boxes, but later switched to a more advanced segmentation model to accurately identify, count, sort by size, and track pallets. To optimize space utilization, we used semantic segmentation to distinguish between merchandise, workers, and forklifts, which allowed us to calculate occupancy within specific zones and provide real-time activity percentages. For truck detection, we created a model that could recognize three scenarios: closed door, open door with a truck, and open door without a truck. We also added a post-processing step to refine predictions, resulting in improved model performance.
We began by running models in a PC development environment to ensure rapid iteration and identify the most optimal models. After completing model development, we ported it into an edge solution.
The challenge was finding the right model-platform combination. While some models performed well on Jetson, others didn't meet Jetson Nano's capacity.
Thus, we carefully balanced model and platform selection for efficient deployment. We finally optimized the models for deployment on NVIDIA Jetson Xavier, converting them to formats like TensorRT or Yolact Edge.
Considering model size and hardware requirements, we deployed our solution on Jetson Xavier devices. These devices were strategically placed across sites, connected to video cameras via RTSP for video streaming. Besides, we implemented an automated reporting system that generated and sent reports at the end of each day to ensure seamless communication with our clients.
To maintain uninterrupted operation and ensure system health, we also developed robust health monitoring modules. These modules monitored the daily reports to verify their successful transmission and promptly triggered alarms in the event of any anomalies or issues, ensuring system reliability.
The models showed a 95% accuracy, even on challenging cases where, for example, pallets were stacked one on top of the other. The successful proof of concepts projects allowed the client to move forward to another stage, presenting a persuasive case for potential fund allocation.
“We value their camaraderie and partnership. They truly feel like an integral part of our team, understanding our needs and challenges. Their expertise in computer vision makes them our first choice for innovative solutions that have a positive impact on our business”