Since the invention of computers, people have sought to make them less bulky, more transportable and increasingly powerful, with the goal that one day machines would operate as intelligently and functionally as human beings. Artificial Intelligence (AI), a technology that was once described as the future of computing, is steadily entering our everyday lives: digital personal assistance, music recommendation services, purchase prediction algorithms, self-driving cars and more.
Applying NLP techniques to efficiently carry out human-to-computer interactions, such as bug triaging, data parsing and speech recognition.
Creating solutions capable of predicting harmful actions on video streams using a hybrid of CNN and RNN neural networks architecture.
Utilizing powerful deep learning tools when training models to develop solutions enhanced with image recognition technology.
Enhancing HMI solutions with intelligent algorithms that alert drivers in case of a lane change and eye closure events.
Gather and enrich customer data or use appropriate third party data sources.
Define standard AI/ML tasks and identify state-of-the-art models to solve them.
Train models with customer data.
Validate model with data subsets.
Enrich customer data and adjust model engineering parameters.
The leading US manufacturer and marketer of fishing tackle boxes, archery equipment, game cameras, protective cases and other gear for hunters
As they began the digitalization of their products to continue to lead their industry, they chose Orion to help build an AI-powered image processing service that sends out alarms if an animal or bird is detected by a trail camera. The service had to recognize certain species of elk and deer.
Solution & Technologies
We deployed the AI component as a cloud service that can be deployed on-premise as it does not depend on third party cloud services, and used the Detectron2 library from Facebook as the main framework to detect animals and birds. FasterRCNN architecture was implemented for recognition tasks. Both bounding box and classification accuracy were used for model assessment, forming Average Precision (AP). Training data were provided by the customer, as well as from open sources including OpenImage dataset, iWildCam dataset and Google images.
