Pseudo-Label Self-Training for Lane Marking Segmentation in Rain and Night Video, with Application to Multi-Violation Traffic Evidence Generation

Lane marking segmentation is the weakest stage in camera based traffic rule enforcement, because public road marking datasets are partially labelled and rarely cover rain or night. This paper presents a self-training pipeline for partially labelled lane data and uses it inside a multi-model violation detection framework for Kathmandu roads. A SegFormer-B3 teacher, fine-tuned on the annotated frames, generates lane masks for the unlabelled rain and night frames of the RLMD dataset, an image level confidence filter at 0.6 discards uncertain masks, and the accepted pseudo-masks are merged with the original annotations, which enlarged the low visibility training set by about 40 percent. A class weighted mean confidence, P_mean, is reported alongside IoU as a stability measure for thin markings. The segmentation stage is combined with YOLOv8 helmet and number plate detectors and a Faster R-CNN rule classifier, and each frame produces a JSON evidence record. On held-out splits scored against the datasets' original labels, PSPNet reaches mIoU 0.87 on CeyMo, SegFormer-B3 reaches 0.82 in rain and 0.79 at night on RLMD, and YOLOv8 reaches mAP50 of 0.93 for helmets and 0.91 for plates. On 8,500 frames from ten Kathmandu videos the integrated output agreed with manual inspection in 87.4 percent of frames. The gain from pseudo-labelling was not isolated in a controlled ablation and no numerical gain is claimed for it. Keywords: traffic violation detection, deep learning, YOLOv8, PSPNet, SegFormer, Faster R-CNN, semantic segmentation, pseudo-labelling, evidence generation, intelligent transportation systems