An Experimental Vision-Based Predictive Framework for Lane Detection and Vehicle Localization Under Low Visibility Conditions During Driving

Reliable lane detection and vehicle localization are essential for advanced driver assistance systems, particularly under low-visibility conditions such as fog, shadows, low illumination, and degraded lane markings. This paper presents a vision-based predictive framework integrating deep convolutional neural networks, classical image processing, and Kalman filter-based state estimation supported by inertial measurement unit (IMU) data. Lane geometry is estimated from monocular camera images using a CNN-based regression model, while a confidence-adaptive predictive filter exploits vehicle motion to compensate for temporary visual degradation. The system is implemented on low-cost onboard hardware consisting of a forward-facing monocular camera, an IMU, a GPS module, and an embedded processing unit installed on a passenger vehicle. Experimental evaluation is conducted through simulation and real-world driving tests on urban highways using approximately ۲۷۷۰۰۰ image frames collected under diverse traffic and visibility conditions. Results demonstrate an average lane detection accuracy of ۸۶.۰%, improving to ۹۲.۷% with image enhancement. Vehicle localization achieves average lateral errors of ۰.۲۴-۰.۳۱ m and remains stable for up to ۳ s during short-term loss of visual information. Real-time performance is achieved at ۳۰ FPS with an average processing time of ۲۶ ms per frame, confirming suitability for practical ITS deployment.

کلیدواژه ها:

Lane Detection ، vehicle localization ، Deep Learning ، Kalman Filtering ، Intelligent Transportation Systems