مهرداد تنهائي نوقابي

Lane detection is one of the fundamental challenges in advanced driver assistance systems an‎d autonomous vehicles. This research was conducted with the aim of designing an‎d eva‎luating a model based on the RESA-34 architecture, along with a post-processing pipeline consisting of bird’s-eye view transformation an‎d third-order polynomial fitting. Initially, the performance of the baseline model was eva‎luated on the TuSimple dataset. The results indicated that the model is capable of detecting lane markings with high accuracy; however, under conditions such as the presence of shadows an‎d road curvature, issues such as lane discontinuities were observed. To improve the model performance, a post-processing pipeline was designed an‎d implemented. This pipeline includes a bird’s-eye view transformation to reduce perspective distortion an‎d third-order polynomial fitting to achieve more accurate reconstruction of lane lines on curved roads. Quantitative results showed that the baseline model achieved an accuracy of 96.82%, a false positive rate of 3.63%, an‎d a false negative rate of 2.48%, with a processing speed of approximately 74.3 frames per second. Nevertheless, in challenging conditions such as shadows an‎d curved road segments, the output lane lines exhibited discontinuities. To address these shortcomings, the proposed post-processing pipeline was applied. Quantitative eva‎luation demonstrated that the final model with post-processing improved the accuracy to 97.11%, reduced the false positive rate to 3.24%, an‎d decreased the false negative rate to 2.19%. Although the processing speed decreased to approximately 61.5 frames per second on an RTX 3050 Ti an‎d 14.8 frames per second on Jetson Nano hardware, it remained above the threshold required for real-time operation. Qualitative analysis on the TuSimple dataset an‎d local road data collected in Tehran showed that the applied post-processing techniques were able to reconstruct lane lines continuously an‎d accurately, effectively resolving the issues of the baseline model under challenging conditions. Real-time experiments on embedded hardware further demonstrated that the proposed model is capable of detecting lane markings an‎d providing driver warnings in scenarios close to lane boundaries under real-world an‎d hardware-constrained environments. This research emphasizes the necessity of post-processing techniques to improve the quality of lane detection model outputs an‎d shows that the proposed methods have practical applicability in advanced driver assistance systems an‎d autonomous driving applications.

تشخيص خط جاده , مجموعه‌داده TuSimple , مدل پايه ResNet-34 , پس‌پردازش

Lane Detection , TuSimple Dataset , ResNet-34 Base Model , Post-processing

Mehrdad Tanhaee

Dr. Ahmad Ayatollahi