School of Science and Technology 科技學院
Computing Programmes 電腦學系
| Programme | Bachelor of Computing with Honours in Internet Technology |
| Supervisor | Prof. Vanessa Ng |
| Areas | Intelligent Applications |
| Year of Completion | 2020 |
| Award | Champion, IEEE (HK) Computational Intelligence Chapter 17th FYP Competition |
The aim of this project is to implement a visually impaired assistant with neural network and image recognition to help visually impaired go outside efficiently and safely. The visually impaired assistant will be similar to a person to accompany the user, using the image recognition technology to describe the nearby environment to the visually impaired in real- time, letting them know what is going on in front of them. The obstacle detector will let them know the obstacles ahead so that they can be prepared early.
To allow visually impaired people to go to their destinations, navigation is also a must, and there are several apps on the Internet providing the navigation function, but the system will make the navigation system more useful to assist the visually impaired to go to the destination.
The main aim of the project is to implement an application to help visually impaired to go outside. The project has also defined several sub-objectives as follows:
Image Caption
To provide a scene description, we have to use image caption technology and we have chosen “The show, Attend and Tell” for our project as an image caption technology since our target user is Hong Kong people.
Figure 1: Principal of Show, Attend and Tell, retrieve from Xu, et al, 2015
Figure 2: Prediction Caption
Speech Recognition
For speech recognition, Alex Graves (2014) proposed a model of recurrent neural networks and CTC layers, as well as RNN-CTC pronunciation and acoustic models for speech recognition (A. Graves and N. Jaitly., 2014).
Object Detection
To provide obstacle prompt, we have to use object detection technology. MobileNetV2 is a lightweight single shot multi-box detector, it is suitable to run at most of the mobile devices and embedded devices. It will be used as a backend of our obstacle prompt system since real-time performance is more important for obstacle prompt systems when the user may not know the obstacle.
Figure 3: Example output of object detection, retrieve from Tensorflow
Depth Estimation
To provide more accurately obstacle prompt, we have to use depth estimation technology to sort out the most dangerous obstacle to the users. High Quality Monocular Depth Estimation via Transfer Learning (DenseDepth) is an encoder-decoder architecture model, which can present a high resolution depth map by inputting a single RGB image.
Figure 4: Original Color Image
Figure 5: Depth map
The system has consisted of three components: Android Client, Raspberry Pi, and Machine Learning Server. The responsibilities of different components are shown as below.
Figure 6: Major Components
Figure 7: Components of the system
Android Client
Raspberry Pi
Machine Learning Server
Figure 8. Open “Visually Impaired Assistant”
Figure 9. Voice Recognition Function
Figure 10. Voice Recognition Function
Figure 11: Voice Navigation screen
Figure 12: Navigation screen
Figure 13: Sense Description screen
The voice recognition interface will be displayed when the user clicks the voice recognition button on the main screen or clicks the voice navigation on the bottom navigation bar. Users can say where they want to go, and the application will automatically open the navigation screen.
The navigation screen shows the route to the destination and the route information will be read out by Text-to-speech service.
The scene description interface shows the image of the external camera and the scene description of the picture, which will be read aloud by Text-to-speech service according to the priority.
Figure 14: Setting screen
Figure 15: A user hold the devices with a walking stick
Figure 16: A user hang around the devices to the neck
The setting screen can set whether to turn on 3D direction prompts and related prompt intervals.
The user can hang the raspberry pi and the external camera (devices) around the neck or hold it with a walking stick. After that, the user needs to put on the earpiece and microphone, and then turn on the app using their mobile phone.
The user can listen to the nearby environment through the earpiece, if there are obstacles in front such as lampposts, stairs, signs, etc., the application will use voice to prompt the user.
In our observations, testers used our equipment and application to collide with objects and people less often than when not in use. The number of object collisions has been reduced by about one-third, and the number of human collisions has dropped significantly by more than one-half. The data indicate that the use of our devices and application can reduce the chance of partial accidents and thus improve user safety. Besides, users can reach the destination more efficiently when using our solution with 3D direction prompts, as can be seen from the time to find the correct direction and the time travel to the destination.
Table 1: User Evaluation Average Result
This project successfully implemented a navigation system that uses the MapBox API and converts its voice into Cantonese to allow the system to work with Hong Kong users. On the basis of the navigation system, a 3D direction prompt system is also implemented, allowing users to distinguish whether they are facing the correct direction through the left-right channels and the large-small beep sound. In addition, the voice recognition system can filter the user’s voice using DialogFlow and output the destination for the Mapbox API to search for the appropriate route.
In the environment description system, real-time environment description is realized, the user can know the affairs in front of the camera through the voice by the text-to-speech service. Environmental description can let the visually impaired know the nearby information that not only be told by other people.
The obstacle prompts system has also been successfully implemented. That system uses image recognition technology to recognize the objects in the image and prompts the user through voice when an obstacle is found.
Future work
It can be improved in hardware and software.
Hardware
Software
