A paced multi-stage block-wise approach for object detection in thermal images

Area of Science:

• Computer Vision
• Machine Learning
• Thermal Imaging

Background:

• Accurate object detection is crucial for thermal imaging applications like autonomous vehicles and surveillance.
• Conventional methods struggle in low-light or nighttime conditions, limiting thermal imagery's utility.
• Deep neural networks trained on natural images offer potential for thermal domain adaptation.

Purpose of the Study:

• To propose a paced multi-stage block-wise framework for effective object detection in thermal images.
• To leverage pre-existing knowledge from natural image datasets for improved thermal domain performance.
• To enhance accuracy and training efficiency in thermal object detection.

Main Methods:

• A multi-stage, block-wise framework designed for thermal image analysis.
• Domain adaptation of deep neural network-based object detectors using a 'pace' parameter.
• Utilizing pre-trained models on large-scale natural image datasets.

Main Results:

• The proposed framework achieved superior performance on the FLIR ADAS dataset for person, bicycle, and car detection.
• Demonstrated enhanced accuracy and training efficiency compared to existing benchmarks.
• Showcased superior performance on night-time images over state-of-the-art RGB detectors.

Conclusions:

• The paced multi-stage framework effectively enhances object detection in thermal imagery.
• The approach shows strong generalizability across different thermal datasets.
• This method provides a robust solution for challenging low-light and nighttime thermal detection scenarios.