Depth-Enhanced Deep Learning Approach For Monocular Camera Based 3D Object Detection

Wang, Chuyao; Aouf, Nabil · 2024 · Crossref

DOI: 10.1007/s10846-024-02128-w

archive: archived pipeline: cataloged verified

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

Summary

This paper addresses the challenge of accurate 3D object detection using monocular cameras for autonomous driving. While LiDAR-based methods provide precise spatial data, they are expensive and computationally demanding. Monocular approaches are cost-effective but suffer from the inherent lack of depth information in single images, leading to significant performance gaps. To bridge this gap, the authors propose a depth-enhanced deep learning framework that integrates depth estimation directly into the detection network to improve spatial perception without excessive computational overhead. The proposed method utilizes the NuScenes dataset and employs a VoVNet-v2 backbone connected to three novel components. First, a Feature Enhancement Pyramid Module (FEPM) replaces standard Feature Pyramid Networks by using an asymmetric fusion mechanism. This module captures contextual correlations across multiple scales by reshaping feature maps and applying self-attention mechanisms to fuse high-level and low-level features, thereby enhancing the representation of large instances and global context. Second, an Auxiliary Dense Depth Estimator (ADDE) is introduced during training to generate dense depth maps. This module updates the backbone and FEPM parameters to improve depth perception capabilities but is removed during inference to maintain efficiency. Third, an Augmented Center Depth Estimation (ACDE) module improves depth accuracy by regressing the depths of the four top vertices of the 3D bounding box. Using geometric constraints, the final center depth is calculated as a confidence-weighted average of the direct center depth prediction and the averaged vertex depths, allowing the model to dynamically select the more reliable estimator. The model is trained using stochastic gradient descent with specific loss functions, including inverse smooth L1 norm for depth tasks and cross-entropy for classification. The training process involves two stages: first, training the full model with the ADDE for 12 epochs, followed by fine-tuning the depth-related regression branches for another 12 epochs. The detection heads predict object classes, dimensions, orientation, velocity, and offsets. Experimental results on the NuScenes benchmark demonstrate that the proposed approach outperforms existing state-of-the-art monocular 3D object detection methods. The integration of the FEPM significantly enhances contextual feature representation, while the ACDE module provides more robust and precise depth predictions by leveraging geometric consistency. The ADDE effectively improves the network's spatial perception during training without adding inference latency. The study concludes that this depth-enhanced architecture offers a promising, lightweight solution for autonomous driving applications, achieving superior detection accuracy compared to competitive methods that rely on independent depth estimators or complex multi-modal frameworks.

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