Artificial Intelligence at NEXYAD


Many publications of the team and of founders : Georges STAMON (pattern Recognition & AI), Jean-Louis AMAT (Artificial Intelligence for image processing), Gérard YAHIAOUI (Deep Learning and Fuzzy Sets), Pierre DA SILVA DIAS (Deep Learning, Driving Risk Computing), etc … NEXYAD is one of the most experiencedprivate team in the world on research and applications of AI.

Many advanced internal tools for AI and for XAI (XAI is the new research programm of DARPA in AI – NEXYAD team is involved in XAI since 1992 on military applications) :
Deep Learning, Fuzzy Knowledge Based Systems, Genetic Algorithms, Genetic Classifiers, Possibility theory, Bayesian networks, …

More that 20 real world customers projects involving AI. Four software components for car mobility involving AI and XAI.

Expertise, Experience, methods, and methodologies to use AI and XAI algorithms together with internal tuning (not only pushing the buttons of a software)



Avanced development tools and skills for training and validation of Deep Learning for Automotive applications


Deep Learning for computer vision : a special NEXYAD AI architecture automatically selects best preprocessing features extraction, and manages neural learning and testing (KPIs and non regression tests). The whole process is triggered by « pushing a button ». The input is a labellized A.G.E.N.D.A. (*) database.
Real time detection and pattern recognition can be replayed on the framework RT-MAPS on PC (asynchronous time stamped signals, data, and videos).
This environment is used to generate ObstaNex and RoadNex releases.

(*) A.G.E.N.D.A. is a methodology published in scientific papers



Automation of validation and KPIs measurement with « ground truth » labels in NEXYAD database for automotive applications


Recorded road scene Ground truth

Tests of non regression
KPIs measurement
Tests are automated using the asynchronous real time framework RT-M



Innovative methodology for Deep Learning : use of « compact databases » for automotive applications

Key issue : how to adapt deep learning to quality enhancement of cameras if you used 1 billion km to train and validate your deep learning ? Build a new 1 billion km database with new enhanced quality sensor ?
NEXYAD answer : Compact Picky database It is the goal of the methogology A.G.E.N.D.A. (Approche Générale des Etudes Neuronales pour le Développement d’Applications), published by NEXYAD team in scientific papers. Pick only significant examples (knowledge based selection).

NEXYAD Deep Learning-based APIs (RoadNex and ObstaNex) can take advantage of camera technical quality enhancement without a big effort to reshape learning database.



Papers references on A.G.E.N.D.A. methodology


« Building a relevant validation database for camera-based ADAS », Gérard Yahiaoui, Pierre Da Silva Dias, Conference Autonomous Vehicle Test & Development, Stuttgart, June 2016

“Methodology for ADAS Validation: The Potential Contribution of Other Scientific Fields Which Have Already Answered the Same Questions”, Gérard Yahiaoui, Pierre Da Silva Dias, CESA congress Dec 2014, Paris, Proc. Springer Verlag

“Caractérisation et implémentation d’un modèle descriptif d’algorithmes de traitement d’images”, J.L. Amat, G. Karavias , G. Yahiaoui, 14eme colloque international sur le traitement du signal et des images, GRETSI, Juan-les-Pins, Septembre 1993

“Un cadre méthodologique dédié à la conception de solutions neuronales : la méthode AGENDA”, G. Yahiaoui, Colloque “les réseaux neuromimétiques et leurs applications”, Neuronîmes92, Nîmes, 1992,



Advanced development tools and skills for building digital maps based mobility applications


We can replay trips with SafetyNex on the real time asynchronous framework RT-MAPS. RT-MAPS feeds SafetyNex with MAP, accelerometers signals, and GPS data, plus ADAS sensors outputs if they were used (case of advanced autonomous driving for instance) with the exact same time stamps than during onboard acquisition.

Everything can be pasted step by step on the map then no bug can happen without being traceable and understandable.
Support for integration into OEMs products is efficient using those tools.



Complex connected architecture skills for mobility applications


Connected objects and complex real time smartphone App skills

Demo smartphone App of SafetyNex API uses :
. Benomad SDK (vectorization of map, next probable road)
. Azure SDK (cloud)
. Stores (Google Play, and Apple Store via TestFlight)
. Facebook SDK (incentive for good drivers)
. Both on Android (in Java) and IOS (in Objectif C)
. SafetyNex SDK/API in native code (C++) : advanced signal processing, and Artificial Intelligence (Fuzzy Sets, Possibility Theory)
. Skills on filtering and undertanding of digital map shapes, dimensions, and ROIs
. Skills on hybrid map (use of tiles uploaded in real time when needed)

All those SDKs are used together in real time (asynchronous threads with time stamps)
CPU consumption of SafetyNex API on an iPhone 5 : 6% of CPU
TRL 9 (Technology Readiness Level)



Team skills on connected cloud security for onboard automotive software Apps


User Licence server, Data recording … use the private NEXYAD cloud on MICROSOFT AZURE in respect with standards of security (OAUTH2, Active Directory, Temporary token, crypted data, …).
And in respect of GDPR (NEXYAD was part of the French Think-Tank on GDPR for Connected Car data privacy Link)

NEXYAD cloud and team skills are important for scalability : user licenses distribution, data storage, etc … can be done automatically with the best level of computer security (already audited by an Insurance Company IT department).