Thesis subject

Deep Reinforcement Learning for Controlling Infrared Sensors and Cameras in Animals Monitoring

Wild animal monitoring is important for preservation of various endangered species. Wildlife monitoring system has progressed with the development of Global Positioning System (GPS) and wireless sensor network (WSN). Meanwhile, there are still many challenges. One of the challenges is the lifetime of devices. Most existing works for animal monitoring use cameras to shoot all the pictures of animals and send them back to the cloud server for further recognition. These approaches consume much energy on capturing images/video. Most devices rely on battery for monitoring animals in wild area. On the one hand, it is difficulty or even impossible to change battery frequently in a wild area. On the other hand, if we increase the size of battery, the size and cost of the devices will increase. Therefore, we need a solution to minimize the power consumption in animal monitoring system.

This project is related the project “Monitor Activity of Wild Animals using Thermopile Infrared Sensors and Cameras”, which includes a two-stage monitoring system using thermopile sensors and cameras. The aim of this project is NOT to set up the system. This project is to control the operation of IR sensor and camera using reinforcement learning.

The accuracy of IR sensor image based animal recognition is low, because of the natural fuzzy property of IR images. Therefore, the question is “how to decide the timing to switch on the camera after the IR image based animal recognition”? If the IR sensor switches on the camera immediately after recognizing an animal in IR image (no matter what kind of animal it is), this may cause false-alarm. There is a chance that the animal captured by the camera is not the target ones, which causes wasting unnecessary battery energy. If the system switch on the camera with 100 percent accuracy of the animal type by recognizing several IR sensor images, then the camera may miss some time to take pictures of the animal.

This project aims to balance the energy consumption of the system and the number of effective animal pictures by using reinforcement learning. Specifically, the student is required to do the following work:

  1. Build a simulation time flow for the behavior of wild animals. The experiment will rely on this simulation environment, which will accelerate the experiment.
  2. Take IR and Camera videos of animals, and allocate the pictures of IR and Camera in the simulation time flow.
  3. Make animal recognition based on IR sensing images and machine learning model.
  4. Use deep reinforcement learning to control the time of switching on camera.


  • Munkhjargal Gochoo ; Tan-Hsu Tan ; Tsedevdorj Batjargal ; Oleg Seredin ; Shih-Chia Huang. Device-Free Non-Privacy Invasive Indoor Human Posture Recognition Using Low-Resolution Infrared Sensor-Based Wireless Sensor Networks and DCNN. 2018 IEEE International Conference on Systems, Man, and Cybernetics (SMC). 7-10 Oct. 2018.
  • Munkhjargal Gochoo ; Tan-Hsu Tan ; Shih-Chia Huang ; Tsedevdorj Batjargal ; Jun-Wei Hsieh. Novel IoT-Based Privacy-Preserving Yoga Posture Recognition System Using Low-Resolution Infrared Sensors and Deep Learning. IEEE Internet of Things Journal ( Volume: 6 , Issue: 4 , Aug. 2019 )
  • Michael A. Tabak Mohammad S. Norouzzadeh David W. Wolfson. Machine learning to classify animal species in camera trap images: Applications in ecology. Methods in Ecology and Evolution, 2019.
  • Van Hasselt, Hado and Guez, Arthur and Silver, David. Deep reinforcement learning with double q-learning. Thirtieth AAAI conference on artificial intelligence. 2016.

Contact person