Multiple object tracking using artificial intelligence
Multiple Object Tracking (MOT) is a key focus in machine vision research. It seeks to follow and label several objects in a video over time, calculating their paths. The challenge is to create a model that works well in tough conditions like occlusions and shadows. Different MOT models based on v...
Saved in:
Main Author: | |
---|---|
Other Authors: | |
Format: | Thesis-Master by Coursework |
Language: | English |
Published: |
Nanyang Technological University
2023
|
Subjects: | |
Online Access: | https://hdl.handle.net/10356/172688 |
Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
Institution: | Nanyang Technological University |
Language: | English |
Summary: | Multiple Object Tracking (MOT) is a key focus in machine vision research. It
seeks to follow and label several objects in a video over time, calculating their
paths. The challenge is to create a model that works well in tough conditions
like occlusions and shadows. Different MOT models based on various principles
have been suggested, and while accuracy has improved, there’s still a need to
enhance the model’s performance.
The dissertation employs BoT-SORT [1] as its baseline. However, BoT-SORT
faces challenges in object recognition and re-identification, particularly in sce narios with object overlap, occlusion, and uniform visual attributes [2] [3] [4].
To address this, the dissertation integrates various re-identification (ReID) back bones with the BoT-SORT architecture to assess resulting model performance.
Among the three models created, the one with the SBS R50 ReID backbone
achieves the highest IDF1 score of 81.9%. To enable visual observation and fa cilitate comparison of tracker performance, tracking samples are executed under
diverse video conditions. The dissertation includes a detailed analysis of perfor mance results and a discussion of future research directions for MOT.
Keywords: Multiple Object Tracking, BoT-SORT, Re-identification. |
---|