Real-time perception systems with low power consumption (T2)

Objectives: Development of solutions for real-time perception systems with low power consumption for use in mobile systems

Research challenges / Novelty / Innovation
• Development of real-time and low-power models: supervised, self-supervised, weakly supervised, and unsupervised
• Distillation and adaptation of Large Vision Models (LVM) for real-time and low-power solutions
• Automatic generation of realistic and annotated synthetic datasets
• (Semi-)automatic generation of annotated real-world datasets
• Data-centric approaches for adapting deep learning models
• Techniques for adapting deep learning models
• Perception and representation of the environment for real-time and low-power systems (2D and 3D) ~ fusion, detection, classification, segmentation, tracking, and 3D reconstruction

Research results:
• Demonstrator: platform for real-time and low-power perception systems

Inovation:
• Innovative learning models, new distillation techniques and optimization of the learning process, (semi-)automatic generation of 2D and 3D annotations, increased adaptability to new environments, optimization of developed models for real-time execution on resource-constrained systems through the use of hybrid architectures (deep learning + classical methods), traceability in portable formats (ONNX, libtorch, TensorRT), and reduction of computational complexity without compromising accuracy and robustness in unseen scenarios