Accurate lane detection is critical for navigation in autonomous vehicles,\nparticularly the active lane which demarcates the single road space that the\nvehicle is currently traveling on. Recent state-of-the-art lane detection\nalgorithms utilize convolutional neural networks (CNNs) to train deep learning\nmodels on popular benchmarks such as TuSimple and CULane. While each of these\nmodels works particularly well on train and test inputs obtained from the same\ndataset, the performance drops significantly on unseen datasets of different\nenvironments. In this paper, we present a real-time robust neural network\noutput enhancement for active lane detection (RONELD) method to identify,\ntrack, and optimize active lanes from deep learning probability map outputs. We\nfirst adaptively extract lane points from the probability map outputs, followed\nby detecting curved and straight lanes before using weighted least squares\nlinear regression on straight lanes to fix broken lane edges resulting from\nfragmentation of edge maps in real images. Lastly, we hypothesize true active\nlanes through tracking preceding frames. Experimental results demonstrate an up\nto two-fold increase in accuracy using RONELD on cross-dataset validation\ntests.\n