MOTorchard

The data collection took place in an outdoor orange orchard situated in Taizhou, China. To capture the necessary information, we deployed a stereo camera(ZED 2i, America) on a transport cart. This setup facilitated the acquisition of crucial data including camera poses, point clouds, and Red Green Blue(RGB) images. To ensure the accuracy and reliability of the dataset, meticulous alignment and calibration processes were taken to align all collected data with synchronized timestamps through.To closely align with practical production requirements, human subjects were selected as the primary targets for detection within the orchard environment in this research. The distribution of obstacles within the orchard and their corresponding motion patterns were classified into distinct categories: cross, cross&return, circle, converge, sim(simulated scenarios that accurately represented real operational conditions).We employed the DarkLabel software for precise MOT annotation in our dataset. The dataset encompasses a diverse range of motion scenarios, consisting of a total of 1630 images. The video was recorded at a frame rate of 15 frames per second, resulting in an overall duration of 108 seconds. Both the RGB images and point cloud data exhibit a resolution of 1280×720 pixels, ensuring high-quality visual and spatial information capture. Additionally, for each individual frame within the dataset, we meticulously documented the camera's relative position and pose compared to its preceding frame. It is important to emphasize that this dataset serves solely for validating our MOT algorithm. Therefore, no explicit division into distinct training and testing sets has been implemented as part of its construction process.