Skip to main content

Hitachi

Corporate InformationResearch & Development

Successful Development of AI Technology for Human Action Recognition in Video under Low Visibility & Slight Variation in Motion

Recognition accuracy for difficult-to- distinguish actions is improved by up to 53% with learning both video and sensor information

March 12, 2020

Hitachi, Ltd. has developed a human action recognition AI technology that can recognize actions in video in which a part of the person is occluded or where there is slight variation in motion that is difficult to distinguish. In this technology, by making AI learn the video and sensor signal obtained from multiple body-worn sensors in advance, it is possible to use only video cameras to capture in real time slight variation in actions without using other sensors. We achieved high accuracy in action recognition in real surveillance videos even when a part of the person’s body was obscured. In comparison with other action recognition technology that normally learns the information only from video, our technology could achieve an improvement of up to 53 percentage points for difficult-to-distinguish actions. It is possible to expand the use of this technology to applications such as detecting suspicious behavior in low visibility and crowds or thickets, as well as in accidents caused by minor collisions between workers and machinery in factories. In the future, Hitachi will apply this technology to video surveillance systems to contribute to a safe and secure society and expand its support for safety operations in factories.

Background and issues Addressed

  • In recent years, human action recognition technology based on video analysis using AI has attracted considerable attention, and its application is spreading in various fields, such as detecting suspicious behavior in public areas, supporting safety activities through motion analysis of factory workers, and preventing major recalls.
  • With conventional technology, installation of sensors is required not only during learning but also when recognizing human actions in real time. Furthermore, incorporating one-dimensional sensor information into neural networks for two-dimensional image information is difficult. As a result, the use of systems that recognized human actions through AI learning based solely on images is typical, and it is difficult to distinguish actions in locations where part of the body is occluded or actions involving slight motion variation.

Developed technologies

  • Attention mechanism that automatically selects valid learning information from multiple types of sensor information
  • Cross-modal learning technology that enables learning through combining different types of information


Fig. 1 Difference in comparison with conventional technology

Features of Dataset Constructed for Evaluation

  • The dataset focuses on 37 types of actions for which there is a strong need for recognition.
  • Data from multiple types of body-worn sensors, environmental sensors, and images was obtained simultaneously, under a variety of conditions (e.g., occluded view).
  • Large-scale dataset for action recognition and detection in the relevant fields


Fig. 2 Sample images of actions from published dataset

Confirmed Results

  • In the results of testing the developed technologies using the dataset, compared to the world’s most advanced technology based on learning with images only, an accuracy improvement of up to 53 percentage points (from 11.2% with conventional technology to 64.51% with the new technology) was obtained for difficult-to-distinguish actions such as opening and closing safes, falling, carrying heavy objects, and using smartphones.

Published Papers, Conferences, Events, etc.

  • The results have already been announced at the International Conference on Computer Vision 2019 (ICCV 2019), which was held in Seoul, South Korea, from October 29 to November 1, 2019.

Details of Developed Technologies


Fig. 3 Details of developed cross-modal recognition technology

1. Attention Mechanism that Automatically Selects Valid Learning Information from Multiple Types of Sensor Information

Which kind of body part sensor information is useful for recognition varies depending on the action that the user wishes to distinguish. When incorporating sensor information into video-based recognition model, if it is learned with indiscriminately as supervision signals from sensor, there is a possibility that sensor information from some body parts which are not useful for action recognition will be learned. In the case of the new technology, an attention mechanism*1 has been developed that enables dynamic selection of sensor information for body parts which are useful for the action to be recognized, making it possible to effectively use action information from sensors in video model learning.

2. Cross-modal Learning Technology that Enables Learning through Combining Different Types of Information

This technology has made it possible to construct AI by combining different types of information, such as sensors and images. This approach applies a method of making student models learn information from teacher models, which is known as knowledge distillation*2. The teacher model is constructed by associating learning information selected automatically from sensor information and ground truth relating to actions in videos. The student model is constructed as a model that recognizes actions based on videos alone, and the outputs of teacher model is learned by student model along with the ground truth of related actions (cross-modal learning*3). As a result of this, the inference abilities of the sensor information-based teacher model, which is sensitive to subtle motion variation and offers robust performance in cases of occlusion, are transferred to the video-based student model. When this technology is applied to video surveillance systems, it will be possible to capture slight motion variation and recognize human actions with a high degree of accuracy even with a system that performs solely on camera videos, without sensors through the use of a student model that has completed learning.

The present research demonstrates the results when video model learns the side information from sensors, while applications in field others than action recognition may also be possible, such as using the technology for action analysis based on sensors only by conversely using videos as the teacher, or performing robust human detection that is effective even when the angle of view varies by learning the mutual image information from different angles. With the aim of further testing this technology, Hitachi has made the large-scale dataset containing sensor and image information that it has constructed publicly available*4. Going forward, Hitachi intends to hold a series of related events and workshops for the purpose of stimulating research in this field and creating an environment that enables collaborative technological development.

*1
Attention mechanism: A method of showing key points relating to a model’s output and focusing attention on them.
*2
Knowledge distillation: A learning method for a given task that efficiently teaches output from a teacher model and intermediate information to a student model.
*3
Cross-modal learning: A method of teaching the interactions between different information modalities (e.g., image and sensor information).
*4
https://mmact19.github.io/2019/

For more information, use the enquiry form below to contact the Research & Development Group, Hitachi, Ltd. Please make sure to include the title of the article.