Multiple object tracking
Feng, Yan (2020)
Feng, Yan
2020
Tietotekniikan DI-tutkinto-ohjelma - Degree Programme in Information Technology, MSc (Tech)
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2020-04-06
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202003242861
https://urn.fi/URN:NBN:fi:tuni-202003242861
Tiivistelmä
Multiple object tracking, a middle-level task, is a critical foundation to support advanced research activities, like pose analysis or motion recognition. In this thesis, the relationship between object detection, single-object tracking, and multiple object tracking was explored and discussed.
On this basis, the Single Shot MultiBox Detector (SSD), SiamMask network, and Detect and Track (D&T) model have been utilized, modified and evaluated.
D&T model is the offline Detection Based Tracking (DBT) network. We had observed the benefits of this correlation loss application via researching on D&T model, and we trained the D&T network on the dataset combination containing the person objects so as to make it useful in reality. In addition, SSD had been applied as the detector with the same tracking methods, during which process the effect of the diverse detectors on the MOT experiments could be figured out. The last experiment was executed on the Single Object Tracking (SOT) model-SiamMask. The original SiamMask network is an online Detection Free Tracking (DFT) network. In order to adapt to the situation of multi-target tracking, it had been modified to initialize multiple target objects with an SSD detector at every specific interval.
Having prepared all the multiple object tracking models, we carried out the evaluation with the MOT17DET dataset. The evaluation metrics for multiple object tracking provided us with a standard view of their performance. In this procedure, we also obtained some helpful knowledge and experience for future MOT improvement.
On this basis, the Single Shot MultiBox Detector (SSD), SiamMask network, and Detect and Track (D&T) model have been utilized, modified and evaluated.
D&T model is the offline Detection Based Tracking (DBT) network. We had observed the benefits of this correlation loss application via researching on D&T model, and we trained the D&T network on the dataset combination containing the person objects so as to make it useful in reality. In addition, SSD had been applied as the detector with the same tracking methods, during which process the effect of the diverse detectors on the MOT experiments could be figured out. The last experiment was executed on the Single Object Tracking (SOT) model-SiamMask. The original SiamMask network is an online Detection Free Tracking (DFT) network. In order to adapt to the situation of multi-target tracking, it had been modified to initialize multiple target objects with an SSD detector at every specific interval.
Having prepared all the multiple object tracking models, we carried out the evaluation with the MOT17DET dataset. The evaluation metrics for multiple object tracking provided us with a standard view of their performance. In this procedure, we also obtained some helpful knowledge and experience for future MOT improvement.