Deep Learning based Fire Recognition for Wildfire Drone Automation

By Robin Yadav
Senior Category (Grades 11-12)
Innovation | Big Data / AI, Engineering and Computer Science

Wildfires are one of the most devastating and harmful natural disasters that occur in Canada every year. Firefighters require the best tools and equipment to fight these wildfires and limit their damage. Currently, drones are becoming an increasingly useful asset to firefighters for wildfire monitoring and assessment.

This project leverages deep learning to create a novel video-based fire detection system to add fire recognition and automation capability to drones. Several state of the art deep learning models were trained and compared. Approximately three thousand images of fire were web scraped or found from smaller datasets and then labeled. Various amounts of data were augmented and different types of augmentations were used to increase the size of the dataset. Additional experimentation was done with the training batch size, confidence threshold and IOU (Intersection over Union) threshold to obtain the greatest mAP (mean Average Precision) for fire detection. Unity Real-Time Development Platform was used to simulate a fire front and to automate drone movement. An automation algorithm was designed to assess recognized fire in video and output future movement. Specifically, the drone was tasked to fly parallel to a stabilized fire front by considering the distribution of fire across an image.

In addition to this algorithm, a DJI Tello drone was automated using Arduino technology to fly between GPS coordinates in the real world and send alerts if fire is detected. The greatest mAP value (@IOU 0.5) was obtained by YOLOv3 at 89.5%.In the simulation, the drone was able to maintain a relatively close proximity to the stabilized fire line and GPS based navigation acts as a failsafe. Data preparation significantly impacts the performance of the model suggesting that certain labeling and augmentation techniques make patterns and features of fire more distinct and recognizable. In the simulation, the drone is able to handle general movement but cannot perform more intricate movements such as making tight turns in succession. Due to the high AP and fast inference speed of the model, this system is viable for real time fire detection. Testing demonstrates that automated drones may have the potential for increasing the efficiency of wildfire monitoring and providing firefighters with critical information.

Related Project


Smart DeTech