Vehicle type recognition in night-time scene is a challenging issue to be resolved due to insufficient luminance, complex lighting environment in night-time and scarcity of public night-time vehicle dataset. Hence, in this paper, we analyse and evaluate the performance of several state-of-the-art model architectures including Faster R-CNN, YOLO and SSD for vehicle detection in night-time scene. Through comparison of evaluation metrics, YOLOv3 with DarkNet-53 achieves the best trade-off between detection accuracy and model architecture complexity, with Average Precision (AP) of 87.43%, recall rate of 91.48% and processing speed of 13.06 FPS with UA-DETRAC validation dataset. In addition, daytime to night-time image augmentation techniques through Neural Style Transfer (NST), conditional GAN (cGAN) and Cycle-Consistent Adversarial Networks (CycleGAN) are implemented to increase the number of night-time images for training dataset by translating the daytime images into night-time scene. Among the three approaches, CycleGAN can generate realistic and natural synthesized night-time images which contribute to improving detection accuracy of the vehicle type recognition model from mAP of 91.81% to 96.47%. Finally, we implement multiple objects tracking technique with Deep SORT algorithm to perform vehicle counting.
Vehicle Types Recognition in Night-Time Scene
Advances in Engineering res
Proceedings of the Multimedia University Engineering ; 2022 ; Cyberjaya and Melaka, Malaysia July 25, 2022 - July 27, 2022
Proceedings of the Multimedia University Engineering Conference (MECON 2022) ; Chapter : 15 ; 139-153
2022-12-23
15 pages
Article/Chapter (Book)
Electronic Resource
English
Recognition distances of vehicle rear markings at night
TIBKAT | 1977
|Study on lane boundary detection in night scene
IEEE | 2009
|Real-time visual recognition and scene interpretation for the intelligent vehicle
Automotive engineering | 2002
|