Deep Learning for Safe Autonomous Driving: Current Challenges and Future Directions

Muhammad, Khan; Ullah, Amin; Lloret, Jaime; Del Ser, Javier; de Albuquerque, Victor Hugo C. · 2020 · OpenAlex-citations

DOI: 10.1109/tits.2020.3032227

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This survey addresses the critical need for reliable deep learning (DL) architectures in autonomous driving (AD), specifically focusing on vehicular safety. Motivated by the high incidence of accidents caused by human error—such as distraction, speeding, and drowsiness—the authors aim to bridge a gap in existing literature by comprehensively analyzing DL methods for seven key safety tasks: road detection, lane detection, vehicle detection, pedestrian detection, drowsiness detection, collision avoidance, and traffic sign detection. The paper organizes these tasks within a three-step pipeline of Measurement, Analysis, and Execution (MAE), evaluating state-of-the-art strategies to highlight their achievements, limitations, and suitability for real-time applications. The methodology involves a critical review of recent DL approaches, categorizing them by their role in the MAE pipeline. The authors assess these methods using standard evaluation metrics, including F-measure, precision, recall, average precision (AP), area under the curve (AUC), and runtime. To provide empirical evidence, the survey benchmarks several models against established datasets, notably the KITTI benchmark for road, lane, pedestrian, and vehicle detection, and the Caltech lanes dataset for lane detection. The analysis compares various architectures, such as Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and hybrid models, examining their performance across easy, moderate, and hard scenarios. Key findings reveal significant trade-offs between accuracy and computational efficiency. In road detection, RBNet achieved the highest F-measure score on the KITTI dataset with a processing time of 0.18 seconds per frame, outperforming other models like DNN and s-FCN-loc, which required longer processing times (up to 2 seconds) despite high accuracy. For lane detection, the DMS method achieved a high AP score of 84.7, while an RNN-based approach achieved an AUC of 0.99, though some VGG16-based models were deemed inefficient for real-time use due to high latency. Pedestrian and vehicle detection results on KITTI showed that while performance is strong in "easy" scenarios (approx. 85% for pedestrians, 93% AUC for vehicles), accuracy drops significantly in "hard" conditions (below 65% for pedestrians, 79% AUC for vehicles), highlighting a major challenge for safety-critical systems. The significance of this work lies in its comprehensive mapping of DL capabilities to specific AD safety requirements, providing a reference for researchers and industry practitioners. The authors conclude that while DL has matured for visual sensing, challenges remain in handling complex environments, ensuring real-time performance, and achieving human-level perception in difficult scenarios. The survey identifies future research directions focused on enhancing the applicability of DL in realistic vehicular environments, emphasizing the need for robust, low-latency models that can reliably support the transition toward higher levels of automation.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-26; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• distraction detection algorithms