Author(s): Xiaolong Wang; Guocheng An; Yanwei Zhang; Fanli Xia

Linked Author(s):

Keywords: Ttention Mechanisms; Deep Learning; Optical Flow Estimation; Recurrent Neural Networks; Video Velocimetry

Abstract: Estimating river surface velocity is crucial for water resource allocation and flood prevention. Traditional intrusive flow measurement cannot monitor the whole river in real-time. In contrast, non-contact image processing technologies can access overall river flow distributions in real-time. However, current optical flow velocimetry methods have poor accuracy at high river speeds. To address this, we propose an improved optical flow velocimetry method. The proposed method consists of three independent modules: optical flow estimation, perspective transformation and flow field estimation. It uses a ReplKNet feature extractor with large convolutional kernels to enhance the receptive field and extract more global features, avoiding omitted extraction at high velocity. The optical flow optimization module incorporates FlowAttention, GRU, and position encoding for better temporal-spatial awareness. After multiple iterations, the model more accurately predicts optical flow. Experiments on the optical flow dataset and the real river dataset show that the method performs well for in large pixel displacements.

DOI: https://doi.org/10.3850/iahr-hic2483430201-254

Year: 2024