CP2A dataset (CARLA Pedestrian Action Anticipation dataset)

在本文中，我们提出了一个新的模拟数据集，用于使用 CARLA 模拟器收集的行人动作预期。
为了生成这个数据集，我们在 Carla 环境中的自我车辆上放置了一个摄像头传感器，并将参数设置为用于记录 PIE 数据集的摄像头（即 1920x1080，110° FOV）。然后，我们计算通过相机视野看到的与自我车辆交互的每个行人的边界框。我们在 CARLA 模拟器中可用的两个城市环境中生成数据：Town02 和 Town03。
模拟行人总数接近 55k，相当于 14M 个边界框样本。每个行人的临界点是他们过马路的第一个点（以防他们最终会过马路）或相反情况下他们路径的最后一个边界框坐标。过马路行为占行人总数的 25%。我们平衡了数据集的训练拆分，以获得相等部分的标记序列交叉/非交叉。我们使用序列翻转来增加少数类（即在我们的案例中的交叉行为），然后对数据集的其余部分进行欠采样。结果是总共有近 50k 个行人序列。
接下来，将行人轨迹序列转换为等长（即 0.5 秒）的观察序列，训练分割有 60% 的重叠。 TTE 长度在 30 到 60 帧之间。它总共产生了近 22 万个观察序列。

In this paper, we propose a novel simulated dataset for pedestrian action anticipation collected using the CARLA Simulator.
To generate this dataset, we mounted a camera sensor on the ego vehicle within the CARLA environment, and configured its parameters to match the camera used for recording the PIE dataset (i.e., 1920×1080 resolution, 110° FOV). We then calculated the bounding boxes for each pedestrian interacting with the ego vehicle that fell within the camera’s field of view. We generated data in two urban environments available in the CARLA Simulator: Town02 and Town03.
The total number of simulated pedestrians is close to 55k, equivalent to 14M bounding box samples. The critical point for each pedestrian is defined as either the first point where they begin to cross the road (in case they eventually cross) or, in the opposite case, the last bounding box coordinate along their path. Crossing behavior accounts for 25% of the total number of pedestrians. We balanced the training split of the dataset to obtain equal proportions of labeled sequences for crossing/non-crossing actions. We employed sequence flipping to augment the minority class (i.e., crossing behavior in our case), followed by undersampling the remaining portion of the dataset. The final result is a total of nearly 50k pedestrian sequences.
Next, we converted pedestrian trajectory sequences into observation sequences of equal duration (i.e., 0.5 seconds), with 60% overlap between training split sequences. The TTE length ranges between 30 and 60 frames. This yields a total of nearly 220,000 observation sequences.

搜集汇总

背景与挑战

背景概述

CP2A数据集是一个基于CARLA模拟器的行人动作预期数据集，包含55k模拟行人和14M边界框样本，重点关注行人过马路行为（占25%）。数据集经过平衡和增强处理，最终包含50k行人序列和22万观察序列，适用于计算机视觉行为预测研究。