Érzékelés/észlelés területén mélytanuló algoritmusok kidolgozása, elsősorban radar és ultrahangos percepció, külső szenzorokkal

Témavezető: Pál Csaba Kőrös
Bosch
email: PalCsaba.Koros@hu.bosch.com

Projekt leírás

One topic from the following:

USS Raw Data Classification This topic aims to extend the existing transducer-based height classification. Possible directions include - object type classification: improves tracing and safety features (e.g. detecting a pedestrian) - height-range classification: allows a more sophisticated approach for higher level layers - road type detection: helps other neural networks with information about the surface Various ML/DL architectures can be employed (ConvNeXt, EfficientNets, MobileNets, different RNNs, etc.). This can be extended with the multi-purpose neural network topic, e.g. developing a unified neural network for height and object classification.

USS Multi-purpose neural network The aim of this topic is to develop a unified neural network that can handle multiple tasks by using a shared backbone and separate heads for each task. By leveraging a common backbone, the neural network can efficiently share computational resources, which may lead to significant reductions in both training and inference times. This shared structure allows the network to extract general features from the input data before branching out into task-specific heads, which can then focus on fine-tuning these features for individual tasks. This approach not only promotes computational efficiency but also encourages the network to learn more robust and generalizable features. The objective is to investigate the range of tasks that can be accomplished by such a neural network and achieve performance levels comparable to current traditional models, while also minimizing runtime.

Radar Parking online / offline calibration - goal:   - determine and improve vehicle to sensor coordinate system offsets (origin and angles) from the sensor data   - offline: define scenes, record them with vehicle, process them in a simulation environment, output the calib values   - online: apply algo in the vehicle during runtime and estimate the values in the target code (optional / stretch goal)

Hallgató