Nowadays, Artificial Intelligence (AI) has penetrated into all aspects of life. For technology workers, the topics of AI, Machine Learning and Data Mining are more interesting than ever.
The activities related to AI, Machine Learning, Data Mining are taking place everywhere. However, it's still a relative new field which brings us a lot of challenges. For a developer, AI demands more advanced and complicated skills and knowledge, compared to traditional programming. So, most of them are studying and researching it.
On the Vietnamese market, we can find quite a few home-grown products with AI elements in it. These products don't show a lot of creativeness or good use of the technology.
In that context, Green Global's AI team was established with only 3 members at the end of 2017. They were fortunate to have the chance to apply AI and embed it in a real life product. This project is all about a combination of Machine Learning and Data Mining. Our customer expects to develop a smart professional management and monitoring system aiming to optimize work efficiency in spaces such as shops, supermarkets, factories, hospitals, hotels, etc. Phase 1 will be user to equip a chain of stores of the second largest retailer in Japan.
Our solution includes AI Cameras with a CPU. We capture the scene and apply calculation models (AI models) to identify and/or classify objects, postures, actions, etc. These snapshot data-sets will be sent to the server, where it will take responsibility for analyzing and aggregating to formulate conclusions, judgments and suggestions for business leaders.
We can skim several basic features as follows:
- • Discriminating staff and customer, recording staff's activities, evaluating work efficiency and arranging work schedules aiming to get the highest efficiency.
• Identifying new and returning customers, finding age, gender and shopping habits to categorize each group of customers a suitable suggestion or advice.
• Evaluate work space/shop design: adjust and arrange items so it's most convenient for staff and customers.
The system model is generalized as follows:
To get the project going, Green Global’s AI developers have collaborated with a partner to assist with research and integration.
First, we collected sample data and used pre-processing technology to create a training set. From there, we started to train Machine Learning models. We had to deal with Terabytes of data to select and standardize on, for the the different Machine Learning’s purposes. We had help, because a lot of this work has been done automatically by Machine Learning Models. The training set is then deployed, to utilize GPUs for the heavy calculations involved and to produce the necessary models. Each model has a different neural network architecture. For example, Inception, ResNet or MobileNet are using one or more Computer Vision algorithm like SSD (Single Shot Detector: https://en.wikipedia.org/wiki/Object_detection), OpenPose, Human Pose Estimation, Facenet. For Machine Learning frameworks, we use TensorFlow and Caffe. After the training, we also do checking to rate accuracy and performance.
During the development stage, we have modified, replaced and tested various neural network architectures a number of times to find the best solution. After establishing the basic models, we packed it into installable applications. A web-based smart application management interface has also been built, which simplifies the process of managing Machine Learning models.
There are a lot of cameras in working place, so identifying the objects’ moving direction across multiple cameras is complicated work. Green Global’s developers have tested this by using technology for Multi-Camera People Tracking: Vectorization and Probabilistic Occupancy Map. Identifying characteristics (age, gender, face, group ...) have to go through the mapping process before analysis to confirm it's the same person that moved and passed different cameras.
At the same time, we develop a Data Mining System to analyze and synthesize data, to make comments or suggestions. A relevant part of the result is sent to the manager, and another part is used for automated processing. The most important thing in AI is the ability to self-learn, self-improve over time. Our machine learning models will be upgraded constantly with the result of new data collected during its own operation.
Green Global’s developers’ ability and dedication have gained the partner's trust and profound attachment. Starting with this small group of 3 people, we are currently discussing the plan to increase to 20 people next year. In addition to our success, this project also gives us lots of valuable experience and confidence to do more challenging AI projects. At the start of 2019, we have 6 potential AI Projects!
Because of the skills gained and what we learned, it's not that difficult to predict that in 2019, we - Green Global’s AI developers - will make a giant step forward.