UE4-based Virtual Dataset of Pedestrians

Background

Due to the limited availability of open-sourced datasets for pedestrians, the development of computer vision algorithms pertaining to pedestrian tracking, detection, and segmentation poses a challenge. To address this issue, I have created a virtual pedestrian dataset platform that offers tailored datasets specifically designed for computer vision tasks involving pedestrians. This platform aims to provide researchers and developers with the necessary resources to advance their work in this field.

Features

Fully-Annotated Data. The dataset is built upon Unreal Engine 4, leveraging its capabilities to offer a comprehensive set of annotations for pedestrians. These annotations include bounding boxes, segmentation masks, and tracking lane data. The annotations are shown in Figure 1-2. (The tracking lane data is not shown here.)

Highly-Costomized Human Behaviors. The dataset further allows for extensive customization of pedestrian behaviors through the utilization of predefined animations. Notably, the dataset accommodates group behaviors, where individuals within a group can be annotated as a single bounding box, as exemplified by the two individuals depicted on the left side of Figure 1. The behavior tree serves as the controlling mechanism for governing pedestrian behaviors, from which users can add or remove behaviors as needed. The behavior tree is shown in Figure 3.

Summary

The virtual dataset has been instrumental in enabling me to gain familiarity with the operational and programming logic of Unreal Engine 4, as well as the built-in C++ API. This invaluable experience has allowed my lab to conduct a series of experiments specifically focused on pedestrians. While the dataset already offers a substantial range of annotated data, it remains flexible and can be extended to incorporate additional forms of data as per future requirements. This adaptability ensures that the dataset can evolve and grow in tandem with the expanding needs of our research endeavors.