Thanisorn Sriudomporn, Elektroteknik

​Titel: Optimized pose tracking on edge device

Examiner: Yasemin Bekiroglu
Industrial supervisor: Alex Darborg, Knightec AB
Opponent name: Nauman Khurshid Haider

The purpose of this master thesis is to evaluate the performance of different models for human pose tracking to serve edge computing purposes. In comparison to cloud computing, edge computing performs the computation on edge devices, which has less computing power than servers. Two tasks are tested on the edge device by four different models: FlowTrack, LightTrack, JointFlow, and OwnTrack. The ability to estimate human pose on sequence images of each model is evaluated by the Multiframe person pose estimation task.

The Multi-person pose tracking task shows how accurate the models are for tracking human pose. Posetrack dataset and Coco dataset are used to train models. The mAP (Mean average precisions) and the MOTA (Multiple object tracking accuracies) are calculated to measure the performance of each model on the custom dataset in multi-frame person pose estimation task and multi-person pose tracking task, respectively. The FPS (Frames per second) is measured to measure the speed of each model. The mAp changes from 41.2% for JointFlow to 81.3% for LightTrack, and the FPS changes from 0.29 for JointFlow to 1.77 for FlowTrack in the multi-frame person pose estimation task. The MOTA changes from 21.5% for JointFlow to 42.5% for FlowTrack. In the multi-person pose tracking task, the fastest model and the slowest model are FlowTrack at 0.73 FPS and JointFlow at 0.29 FPS. The results show that the top-down models outperforms the bottom-up models.

Thanisorn and Yasemin
​Join the seminar by Zoom
Password: 1234
Kategori Studentarbete
Plats: E2 Room 3311 Lunnerummet
Tid: 2022-06-07 13:30
Sluttid: 2022-06-07 14:30

Sidansvarig Publicerad: sö 05 jun 2022.