Research topics for students

Active topics for possible student theses and indivudual/team projects.

We will provide all required hardware and sensors. We also have a lab environment with all lab equipment where you can perform development and experiments.

If you are a bachelor student and feel that you can handle some specific master topic, you can take it and contact us at mairo.leier@taltech.ee for further discussion. 

Topics for BSc students

Automate range hood (also known as kitchen hood or vent hood) working process

(Added 06.09.2020)

  1. Expected functionality
    1. When something is being boiled and steam starts to come out from the pot, range hood starts automatically.
    2. Depending on the amount of steam, it is possible to automatically change the speed of range hood (2-3 steps).
    3. When boiling is finished and level of humidty decreases close to the range hood, it stops automatically.
    4. It is possible to switch range hood to manual and automatic mode.
    5. It is possible to set range hood humidity sensitivity when it starts and stops (be aware – two independent steps named here).
    6. Take into account that system should to react to slow humidity changes but more sudden changes as usual cooking in the kitchen is.
  2. Tasks to solve
    1. Develop hardware to integrate microcontroller and humidity sensor with few buttons and LEDs (for feedback) to a prototype.
    2. Develop software that functionality is descibed above.
    3. Write documentation about the prototype. It must contain hardware architecture, schematic, software block diagram and test results (with the list of tests and conditions).

LIDAR based robot localization and path planning

Plan robot movement in a closed area and localize robot position with LIDAR and plan the path.

  1. Expected functionality
    1. Robot can build the map of entire location autonomously
    2. Robot can estimate its location based on LIDAR information
    3. TBD
  2. Tasks to solve
    1. Use ROS/ROS2 to connect all hardware and software
    2. Visualize SLAM (an example of end result https://msadowski.github.io/hands-on-with-slam_toolbox/)

Smart human presence detector  (taken)

  1. Expected functionality
    1. Sensor should turn on lights only when human enters to the room. If there something else moving (robot cleaner, etc) then lights should keep off.
    2. Develop a 
  2. Tasks to solve
    1. Use Omron D6T type thermal sensor and microcontroller to build a prototype.
    2. Use Omron D6T type thermal sensor to detect human presence in the area of interest.

Topics for MSc students

Garage door opening and closing detection algorithm development 

(Added 07.09.2020)

General description

Develop neural network model that detects garage door operational status. Door is opened by motor lifting a door up and closed lifting door down. It is possible to stop opening or closing at any time moment by pressing the button. To detect door movement a 3-axis accelerometer is used that measures movements with 100Hz frequency.

Examples of abnormal behaviour of a door

  • Movement is not smooth
  • Vibrations during the moving
  • Door stopped too early until got fully open or closed because of
    • One side of the door is blocked
    • Door has hit something during the movement
  • Door moving speed is different (too fast or slow)
    • Someone/something is hanging on the door while door is moving
    • Door has damaged
    • Door has faulty mechanical parts

Datasets

  • We will provide you several datasets that simulate opening and closing cycles.
  • Datasets contain different cycles when door is opening/closing normally or has different faults.
  • Dataset features we see useful to use:
    • door movement speed
    • door movement direction
    • door movement linearity
    • door movement duration
    • door movement starting acceleration until constant speed
    • door movement deceleration until stopped
    • door vibrations (amplitude and frequency)
  1.  

Your tasks

  1. Develop features based on accelerometer data that are listed here or develop your own feature that is not listed here.
  2. Develop a (machine learning) model based on this available data. Your model should have single output to indicate the status of the door.
  3. Convert your model for using it on STM32 microcontroller (STM32F429 Discovery Kit) for demonstration with the integrated accelerometer.
    • Particular development board has also
      • integrated 64Mbit SDRAM
      • 2.4′ touchscreen LCD
      • 3-axis accelerometer.
  4. Validate your model with test data and make calculations that include
    1. Algorithm computational complexity
    2. Amount of RAM this model needs
    3. Model accuracy

 

Camera based vehicle location detection using embedded RISC-V Kendryte K210 platform

(Added 24.09.2020)

General description

This project is running in cooperation with the leading ferry operators. When cargo is being loaded to the vessel then each particular vehicle has its own place where they should be located. To detect whether a vehicle is approaching to the right line and place it needs to be detected, what type of vehicle is passing the camera and where it is heading to.
Our task is to develop a solution to detect (cargo) vehicle positions. When vehicle is approaching or passing the camera, there is a need to detect the exact location of the vehicle on the vessel. To achieve that, you need to know where the camera is located on the vessel to calculate absolute coordinates of the vehicle on the vessel.

Your tasks

  1. Research about methods how to perform vehicle location and direction prediction.
    • It is possible to use either single or dual camera module.
  2. Implement vehicle location and distance estimation algorithm on the selected hardware.
  3. Verify algorithm accuracy, its computational requirements and tune its parameters if needed.
  4. Perform experiments with diferent types of vehicles and camera locations.
Information about the hardware

Camera based vehicle location detection using embedded Jetson Nano hardware platform

(Added 23.09.2020)

General description

This project is running in cooperation with the leading ferry operators. When cargo is being loaded to the vessel then each particular vehicle has its own place where they should be located. To detect whether a vehicle is approaching to the right line and place it needs to be detected, what type of vehicle is passing the camera and where it is heading to.
Our task is to develop a solution to detect (cargo) vehicle positions. When vehicle is approaching or passing the camera, there is a need to detect the exact location of the vehicle on the vessel. To achieve that, you need to know where the camera is located on the vessel to calculate absolute coordinates of the vehicle on the vessel.

Your tasks

  1. Research about methods how to perform vehicle location and direction prediction.
    • Could it be done with single camera, stereo camera, infrared camera(s) or camera combined with other sensor (radar, LIDAR)?
  2. Implement vehicle location and distance estimation algorithm on the selected hardware.
  3. Verify algorithm accuracy, its computational requirements and tune its parameters if needed.
  4. Perform experiments with diferent types of vehicles and camera locations.
Additional reading

Vessel cargo plan generation with machine learning algorithm 

(Added 07.09.2020)

General desription
We have developed smart card deck solution to automate cargo loading and unloading process between the port and vessel during years long R&D project with one of the largest ferry operator in Baltic region. Among other developments we have automated cargo planning on the vessel that has always been done manually. Placing a vehicle to the vessel there are a lot of constraints that are important:

  • Vehicle dimensions, weight, type
  • The order when they arrive to the vessel
  • Priority and VIP classes, dangerous vehicles, etc
  • Ship stability system to keep vessel in balance during the (un)loading

Current vehicle placement algorithm is based on different constraints. There is a need for an algorithm that can learn over time about each departure. There is a dataset about 5000 departures that consists information what kind of vehicle is placed on which card deck an on which location. This data could be used for teaching the algorithm

 Task description

  1. Research about input data, ship constraints, priorities and other data required to take into account for cargoplan generation.
  2. Develop a machine learning model to use for cargo plan generation
  3. Prepare dataset for training
  4. Train algorithm and validate its accuracy.

Develop object classification model to ARM based microcontroller to classify different types of vehicles.

General description

We need a low-cost System on Chip hardware that is capable to classify different types of vehicles. This hardware has usually very limited processing capabilities, limited number of interfaces to connect the camera and may not have any kind of hardware accelerator to speed up neural processing. Our goal is to develop a low cost and energy efficient system that can process video stream from one camera and classify ca 10 different types of vehicles (car, VAN, bus, truck, trailer, motorcycle, bicycle, etc).

We have started research in this field and currently we have a solution that could classify one object at the time. If there are more than one object on in front of the camera then only one object is being detected. This is a limitation and we would like get rid of this limitation.

Task description

  1. Research about currently implemented algorithm and other possible machine learning algorithms that are suitable for using on embedded controller (ARM based STM32 for example). Hardware could also be a multi-core microcontroller. 
  2. Collect and prepare dataset from different sources for vehicle classification.
  3. Develop a method to classify more than one objects simultaneously.
  4. Retrain the model with our own dataset and perform experiments to verify the algorithm accuracy
Additional reading

Facial recognition with embedded hardware

General description

We have a facial recognition model that has been used for some time in smart elevator project that recognizes and learns all new faces automatically. We need to improve the model accuracy and modify it to run it on embedded hardware.

Task description

  1. Find optimal hardware configuration that could handle high resolution camera image and could perform real-time facial recognition.
  2. Take the existing facial recognition algorithm and find methods how to increase its accuracy. Test its accuracy after model tuning.
  3. Implement the algorithm on embedded hardware and optimize the model to speed it up (if needed).

Camera based person tracking and location detection on embedded hardware (taken)

(Added 06.10.2020)

General description

We need to detect the person location in the field of camera and track his or her movement. Location of the person is calculated based on the camera location. Detection of the person an be done with image processing or person detection methods and it is up to the student to decide which way to choose. Model must work on embedded hardware and must therefore be lightweight. Person location tracking should be either in real-time or near real-time. This solution is intended to work indoor situations.

Task description

  1. Research about possible methods how to perform person detection on embedded hardware and implement one of those methods and make experiments to validate the accuracy in different conditions.
  2. Research about possible methods how to map person location based on the camera. Implement one method and validate the accuracy in different conditions.

Camera based person activity classification on embedded hardware (taken)

General description

We need to develop a human tracking and activity classification model that can run on embedded hardware to process all data locally. Camera is used indoor, located at the corner of the room. Goal is to predict activity detection for any time moment.

Task description

  1. Research about person activity classification methods that are suitable for using on embedded hardware
  2. Implement chosen activity classification algorithm on embedded hardware (Jetson Nano or similar)
  3. Perform experiments for different types of activities and calculate accuracy for each activity type and with different camera configurations.
 
Additional materials

Radar and camera based object detection for critical infrastructure (taken)

(Added 15.09.2020)

  1. Research questions:
    1. Research possible radar and camera hardware solutions for object detection.
    2. Develop a model how to perform object detection and classification using both sensors.
    3. Develop software for selected hardware and perform initial tests for experiments.
    4. Perform experiments and analyze results.

PhD research topics

Reinforcement learning on embedded systems

Neural network optimization techniques for embedded systems

Research and develop methods for network size and computational requirements optimization. Applying Singular value decomposition (SVD) to a pretrained CNN model. Network pruning which begins with a pretrained model, then replaces parameters that are below a certain threshold with zeros to form a sparse matrix, and finally performs a few iterations of training on the sparse CNN. Combining Network Pruning with quantization (to 8 bits or less) and huffman encoding to create an approach called Deep Compression.

Distributed machine learning on embedded systems