Supervisors:

• Axel Jantsch
• Matthias Bittner

This work is about comparing the performance of Recurrent Neural Network (RNN) architectures with traditional Convolutional Neural Networks (CNN) used for Time Series Classification on STM Microcontroller (MCU) platforms.

Klick here for further information., opens a file in a new window

Supervisors:

• Axel Jantsch
• Matthias Wess

This bachelor thesis aims to implement an AI-based audio effects device, capable of simulating various effects, such as overdrive, distortion, and compression on an Embedded Device. The research will address the following research questions:

• What are the main driving factors for latency of AI-based digital audio effets
• What hardware and software architecture is suitable for implementing the AI-based real-time audio effects device?

Klick here for further information., opens a file in a new window

Supervisors:

• Maximilian Götzinger

In this work, you would adapt and optimize the already existing anomaly detection model to execute it in real-time on an Nvidia Jetson embedded device. Accordingly, this thesis project consists of the following steps:

• Examine the already existing anomaly detection algorithm
• Adapt and optimize this existing model for real-time inference on an Nvidia Jetson platform
• Train and validate the adapted model, and evaluate it and its output

Klick here for further information., opens a file in a new window

Supervisors:

• Axel Jantsch
• Martin Lechner

This thesis project aims to simultaneously detect faces and estimate their poses using deep learning approaches. This thesis project consists of the following steps:

• Select one of the state-of-the-art face detection and pose estimation CNNs, e.g., MTCNN
• Select a public face data set, e.g., WIDER-FACE
• Train and validate for face detection and pose estimation
• Optimize the model for real-time inference on an embedded platform, e.g., IMX8Plus NPU
• Benchmark the results

Klick here for further information., opens an external URL in a new window

Supervisors:

• Axel Jantsch
• Matthias Wess

This thesis project consists of the following steps:

• Train an adaptive neural network
• Apply power estimation method for network architecture selection (Tools provided by CDL EML)
• Measure the selected networks on the embedded hardware (measurement setup provided by CDL EML)
• Compare the results to networks optimized towards latency and FLOP targets
• Improve the network architecture selection based on the gained insights

Klick here for further information., opens an external URL in a new window

Supervisors:

• Axel Jantsch
• Matthias Wess

In this work, the topic will be to extend a latency estimation model to ARM processors. Further, the latency estimation model shall be extended to estimation model shall be extended to measure power as well. It has already been proved on NVIDIA and intel platforms. Estimation models enables the user to adapt a neural network for a hardware platform without having to execute the actual inference and in that way save valuable development time.

The research problem is to find out how well our estimation model maps to an ARM processor and how well the latency estimation model can be applied to power estimation.

Klick here for further information., opens an external URL in a new window

Supervisors:

• Axel Jantsch
• Maximilian Götzinger

This project investigates enhancing Convolutional Neural Networks (CNNs) for object detection and segmentation. Typically, separate CNNs are used for these tasks, which increases computational load. The goal here is to use a single encoder for both, common in CNNs, and assess any accuracy changes. The steps involve selecting a CNN model (e.g., MobileNet), training it for object detection, freezing the encoder for segmentation, training the decoder, fine-tuning the segmentation architecture, and simultaneous testing of both networks.

Klick here for further information., opens a file in a new window

Supervisors:

• Axel Jantsch
• Maximilian Götzinger

This thesis project aims to assess RL-based methods for predicting train/tram routes, incorporating potential switches. It involves selecting three contemporary RL approaches (Value-based, Policy-based, and DRL), utilizing an open-source train/tram dataset (like RailSem19 or OSDAR2023), training and evaluating these RL methods for route prediction, and optimizing the chosen techniques for use on embedded hardware.

Klick here for further information., opens a file in a new window

Supervisors:

• Axel Jantsch
• Maximilian Götzinger

The focus of this thesis is to assess cutting-edge CNNs for rail-track detection. The project involves several key stages, including selecting a state-of-the-art dataset like RailSem19, training and testing multiple CNN models (such as DeepLabV3, PiNet, and CLRNet) for rail-track segmentation, evaluating their comparative performance, and ultimately optimizing these networks to enable hardware acceleration.

Klick here for further information., opens a file in a new window

Supervisors:

• Axel Jantsch
• Maximilian Götzinger

The goal of this thesis is to combine two cutting-edge CTOS CNNs, each trained for distinct classification tasks, into a single CNN. This fusion aims to maintain a complexity level similar to individual CTOS CNNs, all while minimizing any significant loss in accuracy. The project involves several steps, starting with selecting two state-of-the-art CNNs like YOLO-NAS, pruning them to 25-50%, and fusing them using depth-wise convolution on a channel-wise basis. The subsequent evaluation will assess the relative performance of these combined CNNs against the original ones, followed by optimizing the network to enable hardware acceleration.

Klick here for further information., opens a file in a new window

Supervisors:

• Axel Jantsch
• Matthias Wess

This master thesis project aims to design and implement an AI-based audio effects device, capable of simulating various effects, such as overdrive, distortion, and compression.

Klick here for further information., opens a file in a new window

Supervisors:

• Axel Jantsch
• Matthias Wess

This master thesis project aims to investigate the use of FPGA and Vitis-AI for accelerating semantic segmentation neural networks in the context of disaster-scene analysis. We will participate in the 2023 Low Power Computer Vision Challenge (LPCVC), developing models for semantic segmentation on the NVIDIA Jetson Nano platform and comparing them to the FPGA performance on Xilinx FPGAs. Therefore, we evaluate the impact of adaptive networks and quantization techniques on accuracy, latency, and power consumption. Our study will provide valuable insights into the trade-offs involved in using FPGAs for semantic segmentation in disaster-scene analysis and help to identify the best approaches for different applications.

Klick here for further information., opens a file in a new window

Supervisors:

• Axel Jantsch
• Maximilian Götzinger

This thesis project aims to use a stacking approach to filter rail-track switches (if present) using two weak models and a meta-model. Because embedded systems platforms often have limited computational resources, the developed solution needs to be deployable and executable in real-time on such platforms.

Klick here for further information., opens a file in a new window

Supervisors:

• Axel Jantsch
• Maximilian Götzinger

This thesis project aims to extend an existing monocular depth estimation CNN and integrate prior information to predict scale-aware depth estimation. This developed CNN-based solution shall be deployable and executable in real-time on an embedded systems platform.

Klick here for further information., opens a file in a new window

Supervisors:

• Axel Jantsch
• Daniel Schnöll

This study investigates the possibility of approximating statistical rounding through a modified non-linearity. It aims to determine its viability and implications. The project includes understanding statistics, calculating distributions for linear/convolution layers, integrating distributions into non-linear functions, evaluating computational overhead, handling back-propagation, and practical testing against traditional statistical rounding methods.

Klick here for further information., opens a file in a new window

Supervisors:

• Axel Jantsch
• Matthias Bittner

A significant part of the research, development and final inspection of automotive powertrain systems takes place on test benches. A large number of different test benches can be combined in modern test factories, for example engine, transmission, powertrain, vehicle, exhaust, battery or inverter test benches. Complex plants consist of a large number of subsystems. High forces, torques, voltages or currents are often implemented, sometimes in the presence of flammable gases and liquids. For this reason, there are strict rules and regulations to maintain occupational safety. Traditionally monitoring the workplace safety indicates a regularly and manual visual inspection of the engineer and is a tedious and error prone work. In order to automatically detect safety critical states e.g., leaking liquids, missing/shifted covers grids, loose cables, one can apply modern DNN methodologies for their detection.

Klick here for further information., opens a file in a new window

Supervisors:

• Axel Jantsch
• Martin Lechner

The research problem is to determine analyse if and how well 3D convolutions run on embedded hardware. This thesis project consists of the following steps:

1. Download, setup and reproduce results of the PointPillars model on the KITTI car dataset
2. Setup environment on NVIDIA Jetson Nano, TX2 and Xavier and implement the model there.
3. Evaluate performance and latency of the different networks to determine if these networks on embedded hardware are applicable in autonomous cars

Klick here for further information., opens an external URL in a new window

Supervisors:

• Axel Jantsch
• Martin Lechner

This thesis project consists of the following steps:

1. Select a common state-of-the-art segmentation CNNs, e.g. MobileNet3 along with a decoder
2. Select an RNN unit(s) based on expected performance, e.g. LSTM or GRU
3. Train only the RNN unit(s) using a public dataset
4. Evaluate the original and the updated CNN+RNN for video consistency

Klick here for further information., opens an external URL in a new window

Supervisors:

• Axel Jantsch
• Matthias Wess

This thesis project consists of the following steps:

1. Create a post-training quantization baseline for INT8 for several networks
2. Measure the performance for those networks in comparison to FP32 and FP16 on selected hardware (this part can be done with other participants of the CD Lab)
3. Apply quantization aware training for the select networks to achieve better INT8 accuracy
4. Apply quantization aware training methods to achieve lower bit-width quantization
5. Profile the execution of the DNN on provided hardware Accelerators.

Klick here for further information., opens an external URL in a new window