[Paper Review] RONELDv2: A faster, improved lane tracking method

Zhe Ming Chng, Joseph Mun Hung Lew+1 more|arXiv (Cornell University)|Feb 26, 2022

Autonomous Vehicle Technology and Safety4 citations

TL;DR

RONELDv2 proposes an improved, lightweight lane tracking method that enhances lane detection accuracy and real-time performance by introducing lane point variance analysis, exponentially weighted moving average lane weighting, and lane parameter merging. It achieves up to a two-fold reduction in runtime while improving accuracy across multiple datasets and models, especially under challenging cross-dataset conditions.

ABSTRACT

Lane detection is an integral part of control systems in autonomous vehicles and lane departure warning systems as lanes are a key component of the operating environment for road vehicles. In a previous paper, a robust neural network output enhancement for active lane detection (RONELD) method augmenting deep learning lane detection models to improve active, or ego, lane accuracy performance was presented. This paper extends the work by further investigating the lane tracking methods used to increase robustness of the method to lane changes and different lane dimensions (e.g. lane marking thickness) and proposes an improved, lighter weight lane detection method, RONELDv2. It improves on the previous RONELD method by detecting the lane point variance, merging lanes to find a more accurate set of lane parameters, and using an exponential moving average method to calculate more robust lane weights. Experiments using the proposed improvements show a consistent increase in lane detection accuracy results across different datasets and deep learning models, as well as a decrease in computational complexity observed via an up to two-fold decrease in runtime, which enhances its suitability for real-time use on autonomous vehicles and lane departure warning systems.

Motivation & Objective

• To address the limitations of the original RONELD method in handling lane changes, varying lane marking thickness, and delayed detection of new lanes.
• To improve robustness of lane tracking across diverse driving conditions and datasets with distribution shifts.
• To reduce computational complexity for real-time deployment on edge devices in autonomous vehicles and ADAS.
• To enhance accuracy in cross-dataset validation scenarios where test data differs significantly from training data.
• To develop a lightweight, efficient lane tracking solution that maintains high performance under dynamic and occluded conditions.

Proposed method

• The method computes lane point variance across consecutive frames to assess tracking reliability and detect potential lane changes or occlusions.
• It applies an exponentially weighted moving average (EWMA) to dynamically adjust lane weights based on historical consistency, improving robustness to transient errors.
• Lane parameters from matched lanes are merged when they are within a threshold distance (2σ of lane point variance), refining estimates and improving accuracy.
• The lane matching process uses a distance threshold based on the standard deviation of lane point locations, allowing flexible matching under varying road conditions.
• The system retains unmatched lanes with decaying weights to preserve potential detections, especially in occluded or complex scenes.
• The approach is integrated with existing deep learning models like SCNN and ENet-SAD, enhancing their output without requiring retraining.

Experimental results

Research questions

• RQ1 How can lane tracking be made more robust to lane changes and varying lane marking thickness across different road conditions?
• RQ2 What impact does incorporating lane point variance have on improving detection reliability and accuracy?
• RQ3 How does using an exponentially weighted moving average for lane weighting affect tracking stability and performance?
• RQ4 To what extent does merging lane parameters from matched lanes improve accuracy, especially at higher IoU thresholds?
• RQ5 Can the proposed method achieve significant runtime reduction while maintaining or improving accuracy across diverse datasets?

Key findings

• RONELDv2 achieved a 2.5% increase in accuracy at the 0.5 IoU threshold on the TuSimple dataset compared to the original RONELD method.
• The method reduced runtime by up to two-fold compared to an optimized version of RONELD, enhancing real-time suitability.
• A lane match distance of 2σ provided the optimal balance in performance, yielding the highest accuracy across both CULane and TuSimple datasets.
• The lane merge step improved accuracy by 0.003 at the 0.5 IoU threshold on CULane and 0.011 on TuSimple, with the greatest impact at higher IoU thresholds.
• The optimal α value for the EWMA was found to be 0.5, balancing current frame reliability with historical tracking consistency.
• The method demonstrated consistent performance gains across multiple deep learning models (SCNN and ENet-SAD) and datasets, including cross-dataset validation.