We developed a deep-learning model technology including deblurring, human segmentation, and gait recognition.
Deep learning is a branch of machine learning inspired by the human brain and simulates the learning process of human brains using computer models. It utilizes different neural networks. For this product, the major networks are Convolutional Neural Networks (CNN) and Generative Adversarial Networks (GAN). The training process of the models takes about one week, and the development of the product lasts much longer. For future development, the product will be in advance of simultaneous detecting, development of the product into one whole application for users with out any coding experience to use, and also test different gait recognition models to improve the accuracy further.
The system works by uploading different testing videos and then use the trained model previously to test the similarities between the videos. The videos are compared comparatively and the two with the closest similarity resembles the same people. The video is uploaded by changing the path manually first, and then the videos start automatically comparing by analyze the distance calculated by specific loss functions. The most similar videos have the lowest distance, or differences. Thus, the AI supported matching process to identify the clips with most similarity yields us the result of successful authentication.
For the coding process, three deep learning models including deblurring, human segmentation, and gait recognition were used. Each part utilizes specific deep learning models and the end product is the combination of the three testing the videos with the processing. Then the code is connected to arduino in for executing the code on arduino. The arduino can be replaced by any door lock system in the future for improved applications.
A test set comprises 3 videos in total. In 2 of these videos, the same people can be seen running and walking in different clips, while in the third, a different group of people is shown.
The same two people are constant across the thirty comparison sets, while the third person changes with each set. The number of people that were successfully identified across all thirty video sets is then used to calculate accuracy.
In the 30th test case, the comparison group consists of three videos. Videos 1 and 2 are identical, while videos 2 and 3 feature the same individuals but with differences in their walking or running styles.
