Embedded Files
Since the early 2010s, image processing technologies using deep learning have advanced dramatically. Our research aims to develop practical intelligent sensing technologies based on deep learning.
Specific topics include constructing entrance and exit management systems using skeletal information and developing robust methods for recognizing people and objects in industrial environments.
Estimating Product Assembly Progress Using Instance Segmentation
Estimating Product Assembly Progress Using Instance Segmentation
In recent years, the development of AI technology has accelerated the introduction of IoT into factories. Among the many related issues, automation of progress management is important because it contributes to improved operational efficiency. In this research, we propose a method for estimating the progress of manual product assembly using a fixed-point camera. Specifically, product progress is estimated using object detection known as instance segmentation.
※This topic is a joint research project with a company.
Example Detection Result
(Number in the Upper Left: Progress Step)
Anomaly Detection in Pipes Using Image-Generation AI
Anomaly Detection in Pipes Using Image-Generation AI
In recent years, the number of sewer pipes that have exceeded their service life has been increasing. Worm-type robots for pipe inspection have been developed, but inspection methods have not yet been fully established. In this research, we propose a method that uses image-generation AI to detect abnormalities inside pipes using only normal images.
Experimental Results
Optimal Estimation of Object-Detection Thresholds for Object Tracking Using Deep Learning
Optimal Estimation of Object-Detection Thresholds for Object Tracking Using Deep Learning
In recent years, IoT has increasingly been introduced into factories. Object tracking is used when tracking products. When camera-based object tracking is performed, accurate object detection is essential. However, the optimal object-detection threshold changes depending on the environment and the target object. In this research, we study a method for optimally estimating object-detection thresholds for tracking using deep learning.
Dynamically Setting the Threshold for Determining Object-Detection Results
Because of Rights Restrictions on Factory Videos, the Right Figure Is an Illustrative Image. Successfully Reduced False-Detection Images by 25% and False-Detection Locations by 32%
Construction of an Anomaly Detection Method for Preventing Industrial Accidents
Construction of an Anomaly Detection Method for Preventing Industrial Accidents
The working environment in Japan is becoming increasingly diverse due to the growing number of foreign workers and the aging population combined with a declining birthrate. Automation for preventing industrial accidents has therefore become an urgent issue. We are conducting research to establish an anomaly detection method for preventing industrial accidents by using surveillance cameras in factories. Specifically, we are focusing on three points: extending a generative model called GAN to time-series data; extending explainable AI, in this topic ROI in moving images, to time-series data; and designing a high-quality anomaly detection method using these two approaches.
※This topic is a joint research project with a food manufacturer.
Anomaly Detection Flow
Person Identification Using Deep Learning
Person Identification Using Deep Learning
First, a person is detected from the captured image, and the face is then detected from the candidate person region. The detected face is input into a pre-trained model, which outputs the similarity to each registered person. If the similarity to every registered person is below a threshold, the person is identified as an unknown person. If the similarity exceeds the threshold, the person with the highest similarity is selected as the identification result. When face detection is not possible, the person candidate region is tracked to interpolate the person-identification decision.
Top-image : Flow of Person Identification
(Person Detection ->Face Detection->
Person Identification->Region-of-Interest Tracking and Relabeling)
Bottom-image : Example of Identification
Human Detection Using Subtraction Stereo
Human Detection Using Subtraction Stereo
By using the subtraction stereo method and shadow detection proposed by our laboratory, we can directly extract human regions and acquire their distance information even under shadows in outdoor environments. In addition, stereo measurement, which tends to become unstable, can be restricted only to foreground regions extracted by subtraction stereo. This improves the robustness and speed of measurement.
-関連論文-
梅田和昇，寺林賢司，橋本優希，中西達也，入江耕太, "差分ステレオ－運動領域に注目したステレオ視－の提案," 精密工学会誌，Vol.76, No.1, 2010
Algorithm of Subtraction Stereo
Human Detection Flow
Human Detection Using HOG Features and Subtraction Stereo
Human Detection Using HOG Features and Subtraction Stereo
For people-flow measurement using surveillance cameras, it is important to automatically detect people from camera images. We therefore propose a high-speed and high-accuracy human detection method that uses subtraction stereo together with Joint HOG features, which represent the rough shape of a person by combining multiple HOG features.
Human Detection Using Joint HOG Features
Human Tracking Method
Human Tracking Method
For detecting human behavior and people flow, time-series information about people is required. To obtain this time-series information, we propose methods for recognizing and tracking people in images. In our proposed method, a Kalman filter based on a constant-velocity linear motion model is used to predict a person's position. The predicted position is then compared and associated with the actually measured position so that the same person can be recognized and tracked.
We have also proposed a particle-filter-based method that enables robust tracking even in crowded environments where people overlap in the image. By repeatedly performing prediction, measurement, and association, the system tracks each person and obtains time-series information about that person.
Human Tracking
Disparity Image Subtraction
Disparity Image Subtraction
Background subtraction using color images is vulnerable to illumination changes. We therefore propose disparity image subtraction, which uses disparity images that are less affected by illumination changes.
Human regions are obtained by first acquiring a background disparity image, then subtracting it from an input disparity image, and finally determining whether the foreground region corresponds to a person based on the distance value and the size of the region.
-関連論文-
加藤 貴大, 戸田 哲郎, 増山 岳人, 梅田 和昇, "屋外での使用が可能な視差画像差分を用いた人数推定," 第23回画像センシングシンポジウム (SSII2017), IS2-19, June 2017.
Top-left figure : Input Disparity Image
Top-right figure : Subtracted Image After Noise Removal
Bottom figure : Output Result
People-Flow Measurement Method
People-Flow Measurement Method
The need for people-flow measurement is increasing in various situations, such as guiding people during disasters in public facilities including train stations, and in marketing. We therefore propose a people-flow measurement method using KLT and Voronoi partitioning.
Example of People-Flow Measurement
(Red: Moving to the Right; Green: Moving to the Left)
Page updated
Report abuse