Off-campus SJSU users: To download campus access theses, please use the following link to log into our proxy server with your SJSU library user name and PIN.

Publication Date

Fall 2020

Degree Type

Thesis - Campus Access Only

Degree Name

Master of Science (MS)


Computer Engineering


Kaikai Liu


Computer Vision, Deep Learning, Drones, Fire Detection, UAV

Subject Areas

Robotics; Artificial intelligence


Wildfires are a growing problem in the US and worldwide — in the last decade we witnessed some of the most destructive, costliest, and deadliest incidents in recorded history. Possible solutions include early fire detection and preventative scans of wildlands — tasks that can be efficiently realized with an Unmanned Aerial Vehicle equipped with an appropriate sensing payload. This thesis proposes a vision-based multimodal fire detection system capable of early detection of new wildfires from a drone, as well as aerial surveillance of existing ones. This work discusses the design of the system and its effectiveness, which is evaluated on fire image datasets, as well as the data collected by the system over a real-world 80-acre wildfire. The Deep CNN model used in the RGB pipeline achieves the test accuracy of 0.975 on an internal dataset, and a state-of-the-art 0.958 accuracy on an external general-purpose fire detection dataset. Furthermore, the fused RGB+IR pipeline is shown to increase the image classification accuracy by 1.4pp, and the sensitivity by 6.0pp over the RGB-only pipeline, achieving a final accuracy of 0.958 in a real-world wildfire scenario. All results are obtained with attention to embedded hardware constraints, and the proposed system achieves an end-to-end throughput of 20FPS, while relying exclusively on onboard resources. Lastly, this thesis introduces two auxiliary components: the Aerial Fire Dataset, a large open dataset of aerial imagery from real-world fire incidents, and the Fire Perception Box, a multimodal RGB+IR camera hardware with GPU acceleration; both contributions are free and open source. Overall, the system is capable of fully onboard, vision-based fire detection that produces spatial results which can be utilized to obtain real-time wildfire maps — a technology that is very much needed in fire management.