For the difficulty in balancing the accuracy and speed of the traditional statistical calculation method, this study proposed a real-time statistics method of escalator passenger flow for embedded devices. Firstly, a distortion-free scaling method was proposed to maintain the consistency of information between the test and the training sample to avoid affecting the performance of the detection model. Sencondly, the YOLOv4-tiny detection model was optimized by a dimensionality reduction module and group convolution, and a YOLOv4-tiny-fast network was proposed, which significantly reduces the number of parameters and improves the inference speed while ensuring no loss of passenger detection accuracy. Finally, a matching algorithm combining custom optimization matrix and occlusion processing was proposed to solve the passenger tracking problem with less computational effort. The experiment was conducted with video of escalator entrances and exits in a real environment. The results show that the proposed algorithm achieves an average accuracy of 96.66% in passenger flow statistics on the embedded device platform, and the average detection speed reaches 25 f/s which is superior to existing algorithms.

Human action recognition has received much attention in the field of computer vision because of its important role in public safety. However, when fusing the neighborhood features of multi-scale nodes, existing graph convolutional networks usually adopt a direct summation method, in which the same importance is attached to each feature, so it is difficult to focus on important features and is not conducive to the establishment of optimal nodal relationships. In addition, the two-stream fusion method, which averages the prediction results of different models, ignores the potential data distribution differences and the fusion effect is not good. To this end, this paper proposed a two-stream adaptive attention graph convolutional network for human action recognition. Firstly, a multi-order adjacency matrix that adaptively balances the weights was designed to focus the model on more important domains. Secondly, a multi-scale spatio-temporal self-attention module and a channel attention module were designed to enhance the feature extraction capability of the model. Finally, a two-stream fusion network was proposed to improve the fusion effect by using the data distribution of the two-stream prediction results to determine the fusion coefficients. On the two subdatasets of cross subject and cross view of NTU RGB+D, the recognition accuracy of the algorithm is 92.3% and 97.5%, respectively; while on the Kinetics-Skeleton dataset, it reaches 39.8%, both of which are higher than the existing algorithms, indicating the superiority of the algorithm in human motion recognition.

Object tracking is of great significance in computer vision tasks. Recently, with the development of deep learning, the tracking algorithms based on Siamese networks have been extensively applied because of their excellent capabilities. However, the performance of the existing Siamese network modules degrades significantly when dealing with special situations such as large deformation of the target, low resolution, and complex background. To address these aforementioned issues, this paper proposed a tracking algorithm based on a multi-stream attention Siamese network. This algorithm first constructs super-resolution modules and data enhancement modules, which performs super-resolution and data augmentation on the target templates, respectively, so as to improve the feature characterization ability of the target template. Then, the three backbone networks were used to extract the features of the original target template, the super-resolution target template, and the data augmentation target template, respectively, and their features were fused; simultaneously, the channel attention module and spatial attention module are applied in the backbone network to improve the feature extraction capability. Finally, the fused feature map and the feature map to be searched were input into the region proposal network module to obtain the target tracking information. The experimental results show that the algorithm achieved the precision of 0.919, the success of 0.707 on the OTB100 dataset and the accuracy of 0.642, the robustness of 0.149 on the VOT2018 dataset, with operation speed higher than 20 times per second in real scenarios, demonstrating the excellent tracking performance of the algorithm and excellent robustness in handling various complex scenarios.

In recent years, attention mechanisms have achieved great success in the fields of image classification, object detection and semantic segmentation. However, most existing attention mechanisms can only achieve feature fusion in channel or spatial dimensions, which greatly limits the flexibility of attention mechanisms to change in channel and spatial dimensions and cannot fully utilize feature information. To address this issue, this paper proposes a convolutional neural network attention module based on feature similarity and feature normalization (FSNAM), which can utilize the characteristic information of both channel domain and spatial domain. FSNAM consists of a feature similarity module (FSM) and a feature normalization module (FNM). FSM generates a two-dimension feature similarity weight map using the channel feature information and local spatial feature information of the input feature map, while FNM generates a three-dimension feature normalization weight map using the global spatial feature information of the input feature map. The weight maps generated by FSM and FNM are fused to generate a three-dimension attention weight map to achieve the fusion of channel feature information and spatial feature information. Moreover, to demonstrate the feasibility and effectiveness of FSNAM, ablation experiments are conducted. The results show that, for image classification tasks, FSNAM significantly outperforms other mainstream attention modules in improving the performance of the classification network on CIFAR dataset; for object detection tasks, the object detection network using FSNAM improves the detection accuracy of small and medium-sized objects in VOC dataset by 3.9 and 1.2 points of percentage, respectively; and, for semantic segmentation tasks, FSNAM can significantly improve the performance of HRNet model, and helps to achieve an average pixel accuracy increase of the model on SBD dataset of 0.58 points of percentage.