Deep learning based pedestrian detection in thermal camera images for driver assistance systems

In recent years, with the increase in population density, vehicle numbers and traffic congestion in cities, driving safety has become more important. Drivers' inability to detect pedestrians in low light and adverse weather conditions causes fatal traffic accidents for pedestrians, drivers and passengers. While twice as many traffic accidents with injuries occur during the daytime, it is known that the death rate is 44% higher at night. Although there are many pedestrian detection studies using visible light images in the literature, these systems don't function efficiently in low-light conditions such as night and twilight. Thermal camera technology ensures the effectiveness of this systems when visible spectrum cameras are ineffective. The limitations of pedestrian detection studies utilizing thermal images about safe driving necessitate the development of pedestrian detection systems. This study aims to prevent pedestrian-vehicle traffic accidents in low-light and adverse weather conditions and to protect human life with a detection system designed using thermal images. With this motivation, Histogram of Oriented Gradients descriptor with Support Vector Machines classification method, as well as AlexNet, Darknet-53, and ResNet-101 deep learning networks popularly used for object detection, had been employed for pedestrian detection. Randomly selected thermal images from frequently used FLIR ADAS and TarDAL datasets had been divided into training and test groups. Models had been trained and detections performed. By using the ResNet-101 algorithm, success rates of 94% recall, 98% precision, and 95% F1-score have been achieved. The obtained results validate the effectiveness of the proposed method in this study.