Thermal cameras and image recognition in reindeer recognition for safer driving

Abstract

This Master’s Thesis attempts to solve the problem of reindeer vehicle collisions by creating a neural network image recognition algorithm that detects reindeer from thermal camera images. The motivation for choosing this research problem is that annually there are around 4,000 reindeer vehicle collisions (RVC), where both reindeer and human lives are lost. Property damage occurs as well, and the estimated vehicle costs are around 15-20 million euros every year. Additionally, reindeer owners are compensated for the loss of their livestock for around 2,5 million euros each year. There is also the emotional cost for humans involved, such as loss of feeling safe on the roads. Despite the large costs associated with RVCs, there has been a limited number of attempts to solve this issue.

Thermal cameras were chosen as their performance is not hindered by low lighting conditions or poor weather. No image recognition algorithm trained on thermal camera images of reindeer was identified in the literature review, so this is a novel contribution of this Master’s Thesis. The aim is to be able to integrate this algorithm into a mobile phone application or vehicle’s Advanced Driver-Assistance Systems (ADAS) directly in the future, so that the driver could be warned of reindeer on the road, and the driver or the vehicle itself could take precautionary measures and avoid a collision with the reindeer. A thermal camera dataset taken with FLIR ONE Pro LT thermal camera was also created, which includes thermal camera images of reindeer and Finnish Forest Reindeer. Open source thermal camera datasets of deer and humans from Roboflow were also utilized. An architecture for the future application implementation combining the image recognition model, FLIR thermal camera, and an Android phone application is also presented.

This Master’s Thesis takes a Constructive Research Approach in solving this problem. The research questions were RQ1: “how is it possible to detect reindeer from thermal camera images?”, RQ2: what is the accuracy of the image recognition model with open source data?”, and RQ3: “what is the accuracy of the image recognition model when generalizing to the self-sourced material?”. The answers to these are for RQ1 the algorithm developed can be used for reindeer detection from thermal camera images with high accuracy. RQ2’s answer is 0,96, which is the average test accuracy when both training and testing were done with the open-source Roboflow material. The answer to RQ3 is 0,8, which is the average test accuracy for when the algorithm was trained with Roboflow material and tested with self-sourced material. The algorithm created compares well with other image recognition algorithms used for animal detection.