This paper proposes a monocular camera-based depth estimation method to prevent accidents between moving heavy equipment, such as forklifts, and workers at manufacturing environments. Traditional depth estimation techniques, which rely on ultrasonic sensors, LiDAR, and stereo cameras, face challenges such as high costs, complex installation processes, and maintenance difficulties. As a result, monocular camera-based approaches have gained increasing research attention. In this paper, we introduce an improved Monocular Depth Estimation (MDE) method aimed at enhancing forklift collision avoidance. Our method incorporates a variant of the Split-Transform-Merge (STM) strategy, traditionally used in deep learning for image classification, to improve depth estimation performance. Conventional MDE techniques generally perform two tasks: regression, which inputs a single image into the network to produce depth outputs, and classification, where discrete depth value classes (i.e., bins) are predefined, and the network outputs the probability of each pixel belonging to a specific bin. However, a significant limitation of these methods arises when high-dimensional feature maps are projected into low-dimensional bins, leading to information loss. To address this, we propose a novel strategy that thoroughly analyzes the relationship between feature maps and bins estimation, applying a variant STM strategy to minimize information loss. To validate the performance of the proposed method, we compared it with a previous MDE method, i.e., AdaBins, and confirmed performance improvements across four evaluation metrics. The experimental results not only confirm enhanced performance in all quantitative metrics but also reveal significant improvements in visual quality.


    Access

    Check access

    Check availability in my library

    Order at Subito €


    Export, share and cite



    Title :

    Deep Learning-Based Monocular Depth Estimation Method for Forklift Collision Avoidance


    Additional title:

    Communic.Comp.Inf.Science


    Contributors:

    Conference:

    International Symposium on Intelligent Technology for Future Transportation ; 2024 ; Helsinki, Finland October 20, 2024 - October 22, 2024



    Publication date :

    2025-03-12


    Size :

    8 pages





    Type of media :

    Article/Chapter (Book)


    Type of material :

    Electronic Resource


    Language :

    English




    Collision Avoidance Using Deep Learning-Based Monocular Vision

    Rill, Róbert-Adrian / Faragó, Kinga Bettina | Springer Verlag | 2021

    Free access

    Forklift active collision avoidance control system and method

    NING XIAOBIN / HONG YIHAO / WANG ZHENGHAO et al. | European Patent Office | 2024

    Free access

    Obstacle Avoidance of Monocular Quadrotors with Depth Estimation

    Liu, Qi / Zhang, Zhenghong / Xiong, Mingkang et al. | British Library Conference Proceedings | 2022


    Obstacle Avoidance of Monocular Quadrotors with Depth Estimation

    Liu, Qi / Zhang, Zhenghong / Xiong, Mingkang et al. | Springer Verlag | 2022


    Monocular Depth Estimation of Noncooperative Spacecraft Based on Deep Learning

    Zhao, Erxun / Zhang, Yang / Gao, Jingmin | AIAA | 2023