Improvement of intelligent methods for pedestrian detection in far-infrared radiation images

Paulius Tumas

Doctoral dissertation

Dissertations are not being sold



Each year, over 1.35 million lives are tragically lost on roads, according to The World Health Organization (WHO). Even though the European Union (EU) has the safest roads in the world, 221 people are being killed on roads every day, thousands more are injured or disabled, with long-lasting effects. Each year EU introduces new safety measures in cars, lorries, and buses for advanced driver assistance systems (ADAS) to prevent accidents. One of the primary functions of ADAS systems is pedestrian detection based on intelligent systems. The recent development of convolutional neural network (CNN) based detectors has proven excellent results in object detection. However, not many studies have been performed with a low resolution far-infrared spectrum images. Since CNN based object detection training requires many images, a new FIR domain dataset is introduced captured during severe weather conditions called ZUT-FIR-ADAS (ZUT). This dataset is the second biggest open-access FIR dataset containing Controller Area Network (CAN) bus data synchronized with the FIR images. Then state of the art YOLO (You Only Look Once) detector is modified and trained on this newly introduced dataset, reaching 89.1 mAP (mean Average Precision). However, the dataset and detectors comparison revealed that DNN detectors tend to adapt to specific conditions and features from captured images and do not work accurately when different dataset images are provided. For this reason, ZUT and SCUT (the biggest open access FIR domain dataset) datasets were merged, and two parallel experiments were done. The first experiment aimed to find a training approach and optimize detector structure for speed and performance. The experiment showed that it is possible to increase accuracy by more than five mAP units by retraining the detector on images where the detector fails to detect pedestrians the most. The first experiment also revealed a possibility to minimize detector structure and decrease needed floating point operations by four times without losing accuracy. The second experiment aimed to transfer severe weather features from the ZUT dataset to the SCUT dataset. The experiment revealed that newly generated images increased the accuracy of the detector by 9.38 mAP. The thesis results were published in seven scientific publications – three in peer-reviewed scientific papers, four in conference proceedings. Additionally, the results of the research were presented in seven conferences.

Read electronic version of the book:


Book details

Data sheet

Imprint No:
114 p.
16 other books in the same category:

Follow us on Facebook