SeeDoo
Highlights
- Machine learning model training for successful object recognition
- Object detection and tracking
- Object classification with deep learning
- Computer vision software module for image processing
- Deployment on Jetson Nano and Jetson Xavier
Challenge
The client approached us with a request to develop an on-premises device for detecting individuals and showcasing relevant advertisements on DOOH displays in transportation or outdoor locations. The fundamental idea was for this device to analyze captured images, discern the audience’s average attributes, and enable the presentation of targeted advertising for enhanced effectiveness.
The primary challenge was to deploy the computer vision model on edge devices with limited GPU and RAM memory. Another challenge lied in training the model for successful object recognition.
Do you have a similar challenge?
Contact us and let’s discuss how SoftTeco can help your business grow through the power of innovation and digitization.
Solution
In collaboration with the client’s team, SoftTeco effectively created a computer vision software for the device that enables object detection, tracking and classification by using deep learning. Our developers utilized the Fastdup tool and machine learning technologies to improve data visualization and streamline data analysis.
To address challenges linked to the limited memory and slow data processing, the model was transformed into ONNX and TensorRT formats, ensuring its seamless deployment on edge devices such as Jetson Nano and Jetson Xavier.
Tech Stack
Components
Python
PyTorch
Pandas
NumPy
SciPy
How it works
The Device
The device is a compact on-premises box with a high-resolution camera. Along with DOOH displays, it is installed in desired locations, where the built-in camera captures images of individuals within its coverage area.
Technologies
Using Python and PyTorch, the SoftTeco’s team developed a computer vision software that helps to detect and track individuals within the camera’s field of view. This model analyzes the captured visual data, accurately identifying selected class among all attributes and displays targeted advertisements on nearby screens in accordance with the specified class request. This approach tailors advertising directly to the auditory within the designated location, thereby maximizing its impact.
Results
SoftTeco has effectively created a computer vision device within the specified timeframe, aligning closely with the project’s unique requirements. The client is fully satisfied with the result and we plan to further improve the model in our ongoing collaboration.
