Our main objective is to provide the visual and cognitive abilities of human beings to artificial systems. To this end, Computer Vision methods and algorithms rely on modeling how the human visual cortex interprets images and videos sequences.
Through the identification of the objects in an image, their dimensions and relative position to the camera, Computer Vision is capable of extracting characteristics of the objects or people. Such analysis allows to subsequently identify objects and persons as well as to interpret the actions that occur in the scene.
Relation to current technological trends:
Classic Computer Vision algorithms: Study and development of classical computer vision algorithms in time and frequency domain to develop an easy to use a structured framework of algorithms and image analysis tools, optimized for the deployment of solutions with high scalability.
Deep Neural Networks: The emergence in recent years of high-capacity hardware devices such as GPUs has allowed an unprecedented development of algorithms belonging to the family of artificial neural networks. We specialise in the use of different neural network architectures as they are the algorithms that best model the functioning of perceptual processes in the human visual cortex. This approach to image analysis processes allows us to address tasks such as:
- Classification of objects and identification of context in images.
- Recognition of temporal patterns, human action identification.
- Movement modeling.
- Person recognition through the classification of biometric characteristics.
- Re-identification of objects or people in video sequences.
Security in deep learning systems – The advent of deep learning systems and its increasing exposure to online services in a wide variety of applications (classification of Youtube videos, recommendation of products in online sales platforms, facial recognition for banking or other services, video surveillance systems, etc) raises concerns regarding the security of neural networks. New types of attacks have been detected, like adversarial image which combines actual images with noise pattern that deceive the deep learning system. This kind of attacks are becoming trendy in recent times and poses a real threat to AI technology-driven services. The Computer Vision Research Line gets into the different techniques of attack, exploring their characteristics and researching possible countermeasures and defense strategies.
- Strong proficiency in deep learning classification and prediction methods as well as in feature detection and matching.
- Deep learning models for early diagnosis of Alzheimer disease. Based on neuropsychological tests and the neuro-image acquisition system, the vision algorithm provides a better clinical diagnosis of neurodegenerative disorders by identifying the first cognitive signs images obtained from Positron Emission Tomography (PET).
- Tools to identify and track unauthorized vehicles through UAVs onboard cameras, and detect potential threats by analyzing people gestures near critical areas (ZONeSEC project)
- Facial recognition algorithms with a statistical module to reduce the rate of false positive, minimizing the number of parameters to set up (BODEGA project)
- Tools for the emotion detection and sentiment analysis of people including the heart rate monitoring (PERSONA project)
- Feature matching detection from UAVs onboard cameras to inspect pipeline integrity and anticipate maintenance needs (e-Fly project)
Market foresight reports highlight that Computer Vision technologies will undergo steady growth in the coming years. Moreover, deep learning and Computer Vision are a key product differentiator technology when embedded into existing solutions, and this last technology encompasses within the Artificial Intelligence trends which is one of the main growth pillars of Atos.
Although Computer Vision technology can provide support to a wide variety of markets, we focus on Homeland Security sector customers due to their increasing demands of biometrical identification and tracking of persons. As a result, different implementations of the facial recognition systems were integrated into Atos portfolio.