At CEA LIST, the Embedded Artificial Intelligence Laboratory (LIAE) is developing an embedded perception platform designed to dynamically understand the surrounding world and enable real-time navigation. The design of these modern high-tech systems requires a wide range of skills and experience from software, such as innovative computer architecture models (AI/CNN), to hardware including the integration of multimodal sensors.Robust tracking requires knowledge and understanding of the object: its appearance, its movement and its evolution in time. The main objective of this internship is to propose a model to predict the trajectories of detected objects, namely the direction of movement and the speed. The first prediction methods consist in describing rules to determine the motion patterns. However, some approaches are computationally expensive and not always suitable. The goal is to investigate and implement neural prediction techniques, in particular the combination of a CNN (Convolutional Neural Network) and a RNN (Recurrent Neural Network) may be a good starting point. The CNN network would be used to extract features from an input image. As for the RNN network, its objective would be to process historical data to anticipate the next movement of the target.

In this context, the objectives of the internship are:

·         State of the art on existing solutions for the hybrid CNN-RNN networks,

·         Choice of a relevant approach and characterization through learning step on a dataset,

·         Proposal of a lightweight network adapted to the constraints of embedded systems,

·         Implementation on embedded chip type: NVIDIA Jetson Orin.

The candidate will work closely with the other perception team members. At the end of the internship, it will be possible to continue the research work on this topic during a PhD thesis.

#StageList

The ideal candidate is a 5th year engineering student or M2 level, with knowledge of neural networks, AI programming, embedded systems and strong mathematical background.