Accurate traffic flow prediction information can help traffic managers and drivers make more rational decisions and choices. To make an effective and accurate traffic flow prediction, we need to consider not only the spatio-temporal dependencies between data, but also the temporal correlation between data. However, most existing methods only consider temporal continuity and ignore temporal correlation. In this paper, we propose a multi-modal attention neural network for traffic flow prediction by capturing long-short term sequence correlation (LSTSC). In the model, we employed attention mechanisms to capture the spatio-temporal correlations of the sequences, and the model based on multiple decision forms demonstrated higher accuracy and reliability. The superiority of the model is demonstrated on two datasets, PeMS08 and PeMSD7(M), particularly for long-term predictions.

In recent years, traffic flow forecasting has attracted the great attention of many researchers with increasing traffic congestion in metropolises. As a hot topic in the field of intelligent city computing, traffic flow forecasting plays a vital role, since predicting the changes in traffic flow can timely alleviate traffic congestion and reduce the occurrence of accidents by vehicle scheduling. The most difficult challenges of traffic flow prediction are the temporal feature extraction and the spatial correlation extraction of nodes. At the same time, graph neural networks (GNNs) show an excellent ability in dealing with spatial dependence. Existing works typically make use of graph neural networks (GNNs) and temporal convolutional networks (TCNs) to model spatial and temporal dependencies respectively. However, how to extract as much valid information as possible from nodes is a challenge for GNNs. Therefore, we propose a multi-mode spatial-temporal convolution of mixed hop diffuse ODE (MHODE) for modeling traffic flow prediction. First, we design an adaptive spatial-temporal convolution module that combines Gate TCN and graph convolution to capture more short-term spatial-temporal dependencies and features. Secondly, we design a mixed hop diffuse ordinary differential equation(ODE) spatial-temporal convolution module to capture long-term spatial-temporal dependencies using the receptive field of the mixed hop diffuse ODE. Finally, we design a multi spatial-temporal fusion module to integrate the different spatial-temporal dependencies extracted from two different spatial-temporal convolutions. To capture more spatial-temporal features of traffic flow, we use the multi-mode spatial-temporal fusion module to get more abundant traffic features by considering short-term and long-term two different features. The experimental results on two public traffic datasets (METR-LA and PEMS-BAY) demonstrate that our proposed algorithm performs better than the existing methods in most of cases.

Existing pavement crack inspection methods heavily rely on a large amount of annotated samples and require overburdened computational power, which is not affordable on edge devices. Recent methods mostly focus on spatial features, which cannot effectively capture the long, continuous, tender, and thin road features. The masked image modeling (MIM) [1] is an effective way to rebuild the crack primitive features by masking strategy on unlabeled data to reduce the dependence on annotated data, and the convolution on frequency domain provides a flexible and efficient path to capture continuous and tender road features. Inspired by the thoughts of masked frequency modeling, we proposed a masked DCT-domain modeling strategy, named CrackMDM, for efficient pavement crack segmentation. Specifically, we propose a DCT-domain masked modeling method in the CrackMDM model which combines the advantages of separable convolutions and spectral convolutions in the DCT-domain to extract continuous and tender crack structures. Additionally, we introduce the self-supervised pre-training with a masking strategy in DCT-domain using unlabeled crack samples to build crack primitives and to fine-tune the encoder and decoder parameters on labeled crack data to refine crack features in the fine-tuning phase. The CrackMDM model is evaluated on three public benchmarks: CFD, YCD, and GAPs, and achieves state-of-the-art (SOTA) performance with superior inference speed. Codes are available at https://github.com/Jyuan357/CrackMDM.