A carregar...
Publicação
Designing and Implementing a Distributed Prototype for Object Tracking Across Multiple Cameras
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 14.28 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
This report details the work that was concluded in the scope of a project at Symbic GmbH.
The company offers a variety of software solutions for logistics environments, such as warehouses
or truck yards. The project is motivated by the desire to gain insights into the potential for leveraging Multi-Camera Multi-Object Tracking to enhance these solutions. To do this, the goal was
the design, implementation and subsequent evaluation of a distributed framework for the problem.
Vehicles were chosen as the target object class. The pipeline was to be as flexible as possible
and capable of handling live videos in real-time. Taking this into account, the design is modular,
separating the problem into four separate subproblem steps. In the first step, an object detector
finds the bounding boxes of all candidate objects. After this, a re-identification feature extractor
generates feature description vectors for each detection. With the help of these feature vectors, a
single-camera tracker joins the detections into trajectories inside a single camera’s view. Finally,
the multi-camera tracker matches these single-camera tracks into global ones spanning multiple
cameras. This pipeline was evaluated on a benchmark dataset and its performance was compared
to another state-of-the-art algorithm. The results suggest that the pipeline is more or less up to
par with the comparison in favorable conditions. However, they also revealed weaknesses when
dealing with situations with large fields of view, in low ambient light or with a high number of
occlusions. Additionally, the pipeline does not reach runtimes that come close to online performance. The report offers probable causes for this and suggests different ways of remedying these
problems.
Descrição
Trabalho de projeto de mestrado, Informática , 2024, Universidade de Lisboa, Faculdade de Ciências
Palavras-chave
Rastreamento de Múltiplos Objetos com Múltiplas Câmeras Visão Computacional Software Distribuído Trabalhos de projeto de mestrado - 2024