Week 5 – /Purdue CNIT 581 Spring 2022 Project – Update – April 21st

CNIT 581-SDR, Spring 2022.

Project: WS-Bot.

Ruiqi, Ola, and Yi.

Robot Navigation system

We have successfully built a map and point-to-point navigation via Jackal using gmapping ROS package provided by the Clearpath company (Fig.1). The basic navigation part is done, and the CV based navigation system, used for finding people is built, but some parameters need to be tuned. For now, the accuracy is not so high in real-time.

Fig. 1 Gmapping in Rviz

Computer Vision System

We tested the trained YOLOv5 model statically and in real-time (Fig. 2, 3). The accuracy in static recognition is 100% in 50 tests. However, the real-time recognition can not achieve 100%, sometime it will lose the subject, to solve this problem, we added more samples that can be undetected in real-time.

Fig.3 Static human recognition

Fig.4 Real-time human recognition

Wearable System

To collect data, we wear the E4 wristband to act normal behaviour and falling behaviour (Fig. 5). At last, we collect about 400M CSV files. The data we collected includes two types: the change of acc in the z-axis and g-value, and the normal behaviour is labelled as 0 while falling down is labelled 1 (Fig. 6). Then we utilized SVM to classify the data (Fig. 7). Now we are testing the model.

Fig. 5 Collecting behavioural data

Fig. 6 Shows a change in value when the user falls

Fig. 7: Traning result from SVM