Thermal images play a crucial role in object detection for various applications, including surveillance, security, industrial automation, and vehicle navigation, as they can capture thermal signatures. However, challenges such as low contrast, fluctuating thermal patterns, and the absence of specific thermal datasets require specialized algorithms to accurately identify the object and its location. This research explores the field of object detection in thermal images, which is a vital aspect for applications such as surveillance and autonomous systems.
The images used are from Kaggle’s website, FLIR ADAS. We used it in two different ways for classification and segmentation models. This thesis used the Contrast Limited Adaptive Histogram Equalization (CLAHE) approach in image processing to improve contrast. This pre-processing phase attempts to enhance the performance of subsequent object recognition algorithms by addressing contrast related issues in thermal images
The first section of this study is to define a convolutional neural network model that can be used in classification situations. Using 2,956 thermal images, the study evaluates the accuracy, precision, F1 score, and recall of the MobileNetV2 and VGG19 architectures.
The results showed us that MobileNetV2 had an accuracy of 97% and F1 score of 96.8%, while VGG19 had an accuracy of 97% and an F1 score of 96.6%.
This study provides valuable insights into the application of transfer learning methods for thermal image classification and provides a comprehensive overview of the effectiveness of different algorithms in meeting the challenges posed by thermal imaging conditions.
In second section of this study, we take advantage of the YOLOv8 approach, which has been widely recognized for accurate real-time object detection, enhancing detection performance in the challenging region of thermal images. The technique was trained using a selected dataset contain 1898 images, of which 1,443 were used for training, 380 for validation, and 75 for testing. We used this approach to segment objects in thermal images. The average accuracy (mAP50) of 82% of the algorithm shows how effective it is. These results confirm that YOLOv8 is suitable for thermal imaging applications, which are new areas that require accurate object detection under extremely hot conditions.