Person detection and tracking using artificial intelligence
Person detection and tracking are important issues in multiple object tracking. Our dissertation aims to develop efficient models that work in crowded conditions for practical purposes. Two major challenges for these models are getting correct trajectories with pedestrian occlusion and spending less...
Saved in:
Main Author: | |
---|---|
Other Authors: | |
Format: | Thesis-Master by Coursework |
Language: | English |
Published: |
Nanyang Technological University
2022
|
Subjects: | |
Online Access: | https://hdl.handle.net/10356/159396 |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Institution: | Nanyang Technological University |
Language: | English |
Summary: | Person detection and tracking are important issues in multiple object tracking. Our dissertation aims to develop efficient models that work in crowded conditions for practical purposes. Two major challenges for these models are getting correct trajectories with pedestrian occlusion and spending less time without reducing accuracy. We attempt to balance the accuracy and run time in crowded conditions.
We review algorithms for object detection and tracking. The one-stage methods (such as YOLO and SSD) are currently the most popular for object detection because they can achieve high accuracy quickly. As for object tracking, the baseline algorithms such as DeepSORT are widely adopted for practical reasons. Therefore, a model consisting of YOLOv5 and DeepSORT is proposed.
The proposed model is evaluated on the challenging MOT benchmarks. We conduct experiments with different detection models and conclude that accurate detection improves tracking performance. Moreover, experimental results show that the proposed method performs better than two state-of-art models, i.e., UMA and JDE, in crowded conditions (MOT20), achieving 3.740% and 14.065% improvement on MOTA respectively. |
---|