Jinn-Liang Liu

Der-Hau Lee, Jinn-Liang Liu

We propose an end-to-end driving model that integrates a multi-task UNet (MTUNet) architecture and control algorithms in a pipeline of data flow from a front camera through this model to driving decisions. It provides quantitative measures to evaluate the holistic, dynamic, and real-time performance of end-to-end driving systems and thus the safety and interpretability of MTUNet. The architecture consists of one segmentation, one regression, and two classification tasks for lane segmentation, path prediction, and vehicle controls. We present three variants of the architecture having different complexities, compare them on different tasks in four static measures for both single and multiple tasks, and then identify the best one by two additional dynamic measures in real-time simulation. Our results show that the performance of the proposed supervised learning model is comparable to that of a reinforcement learning model on curvy roads for the same task, which is not end-to-end but multi-module.

Der-Hau Lee, Jinn-Liang Liu

Inspired by the UNet architecture of semantic image segmentation, we propose a lightweight UNet using depthwise separable convolutions (DSUNet) for end-to-end learning of lane detection and path prediction (PP) in autonomous driving. We also design and integrate a PP algorithm with convolutional neural network (CNN) to form a simulation model (CNN-PP) that can be used to assess CNN's performance qualitatively, quantitatively, and dynamically in a host agent car driving along with other agents all in a real-time autonomous manner. DSUNet is 5.16x lighter in model size and 1.61x faster in inference than UNet. DSUNet-PP outperforms UNet-PP in mean average errors of predicted curvature and lateral offset for path planning in dynamic simulation. DSUNet-PP outperforms a modified UNet in lateral error, which is tested in a real car on real road. These results show that DSUNet is efficient and effective for lane detection and path prediction in autonomous driving.

Der-Hau Lee, Kuan-Lin Chen, Kuan-Han Liou et al.

Based on the direct perception paradigm of autonomous driving, we investigate and modify the CNNs (convolutional neural networks) AlexNet and GoogLeNet that map an input image to few perception indicators (heading angle, distances to preceding cars, and distance to road centerline) for estimating driving affordances in highway traffic. We also design a controller with these indicators and the short-range sensor information of TORCS (the open racing car simulator) for driving simulated cars to avoid collisions. We collect a set of images from a TORCS camera in various driving scenarios, train these CNNs using the dataset, test them in unseen traffics, and find that they perform better than earlier algorithms and controllers in terms of training efficiency and driving stability. Source code and data are available on our website.