Elektrik ve Bilgisayar Mühendisliği Ana Bilim Dalı Tez Koleksiyonu

Permanent URI for this collectionhttps://hdl.handle.net/20.500.12573/418

Browse

Browsing Elektrik ve Bilgisayar Mühendisliği Ana Bilim Dalı Tez Koleksiyonu by Author "Ağırman, Ahmet Kerim"

Filter results by typing the first few letters
Now showing 1 - 1 of 1
  • Thumbnail Image
    doctoralthesis.listelement.badge
    Nighttime fire detection from video
    (Abdullah Gül Üniversitesi, Fen Bilimleri Enstitüsü, 2022) Ağırman, Ahmet Kerim; AGÜ, Fen Bilimleri Enstitüsü, Elektrik ve Bilgisayar Mühendisliği Ana Bilim Dalı
    With the recent advancements in the field of Computer Vision, the central tasks such as object detection, segmentation or object tracking methods attain all-time high accuracies in natural image sets such as ImageNet, COCO, etc. However, due to the innate downsides of digital images acquired in insufficiently illuminated environments, the conventional methods suffer severely. This specific problem remains unsolved. Especially if the environment is pitch dark and the object of interest is emitting light, the dynamic range of the current digital cameras falls short in this situation and the generated digital image contains almost no perceptible visual texture. One prominent example of this is nighttime forest fire videos. In this thesis, detection of nighttime forest fires from video is addressed as an application of the challenging task, scene perception in low light conditions. The first contribution of this dissertation is developing a novel object tracking algorithm for glowing object in the dark environments. The algorithm allows to track fire and nonfire objects throughout the video. The second contribution of the thesis is proposal of new handcrafted features which are designed to capture spatio-temporal behavior of the glowing objects since there is little or no visual textures to be processed. The results showed that the features are descriptive enough to distinguish fire from the other deceptive light sources. The third contribution is employing deep learning models to automatically extract spatial features with CNNs, and temporal features from bidirectional Long Short-Term Memory (BLSTM) networks. The empirical test results show that a CNN + BLSTM pipeline can effectively detect fires at night with a high accuracy. Finally, a new comprehensive nighttime fire video dataset comprising 1358 positive videos and 334535 of fire frames is created.