Excavator joint node-based pose estimation using lightweight fully convolutional network

Abstract

Current deep-learning-based excavator pose estimation methods usually face problems such as high memory consumption and low operation speed owing to large parameter redundancy. This paper presents a joint node-based excavator pose estimation approach using a lightweight fully convolutional network (FCN) that achieves higher accuracy with lower computation and storage requirements. The method directly encodes excavator joint nodes into multilevel features, and employs a deconvolution head to decode them into heat maps to provide joint node coordinates. The lightweight design is made at two levels: block level (employing depth-wise separable convolution instead of conventional convolution for efficiency) and layer level (employing the slimming technique to optimize layer channels for redundant depth removal). Using images collected from real construction sites, the superiority of the method was validated by comparing it with other state-of-the-art algorithms using various hardware platforms. The results indicate the high potential of excavator pose estimation for edge device deployment.