Pedestrian Re-recognition Based on Memory Network and Graph Structure.

Computer interaction and public safety have great research significance and practical value. Because of the problem that the recurrent neural network used in the existing literature will produce gradient disappearance and gradient explosion when the video sequence is long. Then, we propose a Person Re-identification Network with ConvLSTM (CLPRN) network based on convolutional long and short-term memory networks to solve the short-term memory problem. And then, aiming at the problem of information fusion between frames, we propose a Person Re-identification Network with Graph Convolution (GCPRN) network based on the graph structure, introduce a multi-header attention mechanism, and measure the relationship between frames. The experimental results shows that the Rank 1 of the GCPRN network on iLIDS Video re-identification (iLIDS-VID) dataset reaches $$70.58\%$$ and Rank 5 reached $$81.20\%$$ , surpassing the Unsupervised Tracklet Association Learning (UTAL) and Temporal Knowledge Propagation (TKP) algorithm that reached a high level on the iLIDS-VID dataset.