Anchor3DLane: Learning to Regress 3D Anchors for Monocular 3D Lane Detection (2301.02371v2)

Abstract: Monocular 3D lane detection is a challenging task due to its lack of depth information. A popular solution is to first transform the front-viewed (FV) images or features into the bird-eye-view (BEV) space with inverse perspective mapping (IPM) and detect lanes from BEV features. However, the reliance of IPM on flat ground assumption and loss of context information make it inaccurate to restore 3D information from BEV representations. An attempt has been made to get rid of BEV and predict 3D lanes from FV representations directly, while it still underperforms other BEV-based methods given its lack of structured representation for 3D lanes. In this paper, we define 3D lane anchors in the 3D space and propose a BEV-free method named Anchor3DLane to predict 3D lanes directly from FV representations. 3D lane anchors are projected to the FV features to extract their features which contain both good structural and context information to make accurate predictions. In addition, we also develop a global optimization method that makes use of the equal-width property between lanes to reduce the lateral error of predictions. Extensive experiments on three popular 3D lane detection benchmarks show that our Anchor3DLane outperforms previous BEV-based methods and achieves state-of-the-art performances. The code is available at: https://github.com/tusen-ai/Anchor3DLane.

• The paper introduces 3D lane anchors that bypass BEV transformation, preserving spatial context for precise lane prediction.
• It employs an iterative regression process with bilinear sampling informed by camera parameters to adapt to complex terrains.
• Experimental results on ApolloSim, OpenLane, and ONCE-3DLanes demonstrate superior AP and error reduction compared to BEV-based methods.

Overview of Anchor3DLane: A BEV-Free Approach for Monocular 3D Lane Detection

The paper introduces a novel approach titled Anchor3DLane for monocular 3D lane detection, designed to directly predict 3D lanes from front-view (FV) representations without the need for bird-eye-view (BEV) transformations. The key innovation in this paper is the definition and use of 3D lane anchors that operate in the 3D space, which are projected onto FV features to yield accurate predictions by maintaining crucial structural and contextual information.

Methodological Innovation

Traditionally, BEV-based methods employ inverse perspective mapping (IPM) to warp FV images into BEV, simplifying the lane detection task into a 2D format. However, this process is encumbered by assumptions such as flat ground, which are not universally true, thus often resulting in compromised depth information and accuracy. Moreover, the loss of contextual information due to IPM leads to distorted road objects, which further aggravates inaccuracies.

Contrasting these existing approaches, Anchor3DLane proposes a system where 3D lane anchors are directly defined in the spatial coordinates system. By skipping the BEV transformation, the system alleviates reliance on the flat ground assumption and retains richer context for feature extraction. The anchors are characterized by pitches and yaws, allowing flexibility in fitting lane contours in complex terrains such as uphill or downhill paths.

To achieve this, the authors propose an iterative regression process that samples anchor features directly from FV representations using bilinear sampling informed by camera parameters. This approach not only preserves necessary image context but also iteratively refines the anchors to enhance prediction precision.

Performance Analysis and Results

The experimental results, evaluated on three prominent benchmarks—ApolloSim, OpenLane, and ONCE-3DLanes—demonstrate notable improvements over existing BEV-dependent methodologies. Anchor3DLane achieves superior performance on metrics such as Average Precision (AP), F1 score, and lateral (x) as well as height (z) errors. For instance, on the ApolloSim dataset, Anchor3DLane exhibits an AP of 97.2% on balanced scenes, surpassing previous leading methods.

Significant improvements are especially evident in vehicle maneuvers requiring high accuracy, such as handling rare terrains and diversified visual conditions. The ability to manage elevation changes and maintain prediction consistency underlines the advantage of abandoning the BEV dependence. Furthermore, the paper introduces a global optimization step leveraging the inherent parallelism of lane markings to further refine predictions, effectively reducing error margins particularly in scenarios dystopic of lane divergence.

Implications and Future Directions

Anchor3DLane's ability to operate independently of BEV transformations unlocks new potential in developing monocular lane detection systems that can robustly handle a myriad of road conditions without extensive preprocessing. The architecture is hardware-efficient, requiring comparatively lower computational resources while enhancing prediction accuracy, making it a compelling choice for deployment in real-time applications in autonomous vehicles.

This research prompts further investigation into applications of 3D anchor concepts beyond lane detection to other areas of 3D vision tasks like object detection and vehicle trajectory prediction in autonomous driving. Additionally, exploring the scalability of this approach to systems with multiple view integrations or sensor modalities merits attention for further extending the envelope of autonomous perception capabilities.

In conclusion, the paper on Anchor3DLane presents a clear advancement in the field of monocular 3D lane detection by addressing and overcoming the limitations posed by BEV methods. The proposed solution is not only technically sound but also practically significant, boasting potential widespread application in the rapidly evolving field of autonomous vehicle navigation and beyond.