Dynamic Footprint Retrieval Method Based on Multi-class Feature Fusion

Footprint features, as one of the biological features of the human body, play an important role in the field of personal identification. At present, most research on footprint recognition focuses on footprint images as experimental data, using deep learning algorithms as the foundation and relying on auxiliary algorithms to complete high-precision footprint recognition tasks. However, there is a problem with models built on footprint images. Due to the similarity of footprints of different people, as the number of samples increases, the differences between the features of footprints of different people will continue to decrease, leading to an increasing false detection rate of the model. In order to reduce the interference of similarity between footprints on model recognition ability, this paper takes dynamic footprints as the research object and proposes a dynamic footprint retrieval method based on multi-class feature fusion. The proposed method uses a spatio-temporal fusion module to integrate the spatio-temporal information of footprints, so that the footprint recognition method is not limited to the apparent information of footprints. Firstly, the convolutional neural network is used to extract the frame level features of dynamic footsteps, and then the feature fusion module calculates the complete apparent features of the fused dynamic footprints through a trainable weight matrix and frame level features. Secondly, the temporal aggregation branch of the spatio-temporal feature fusion module is used to extract long-term temporal features within frame level features, and then the long-term temporal features are fused with frame level features through orthogonal fusion calculation method to form spatio-temporal features. Finally, the visual features and spatio-temporal features are fused for dynamic footprint retrieval. A comparative experiment is conducted on a dynamic footprint dataset of 200 people with existing deep learning algorithms, and the experimental results shows that this method achieved better performance, with Rank1 and mAP being 85.39% and 55.28%, respectively.