The Fall Detection Dataset

该数据集是基于事件相机捕获的，用于事件驱动对象检测任务的基准数据集。数据集通过事件相机模拟器将基于帧的跌倒检测数据集转换为事件驱动的跌倒检测数据集，确保面部隐私保护并降低内存使用。该数据集适合于自动驾驶等需要高动态范围、低延迟和低功耗的领域。

This is a benchmark dataset for event-driven object detection tasks, captured using event cameras. The dataset is generated by converting a frame-based fall detection dataset into an event-driven fall detection dataset via an event camera simulator, which ensures facial privacy protection and reduces memory usage. This dataset is applicable to domains such as autonomous driving that demand high dynamic range, low latency and low power consumption.

搜集汇总

数据集介绍

构建方式

The Fall Detection Dataset is meticulously constructed by converting frame-based fall detection videos into event stream data using an event camera simulator. The original dataset, Le2i fall detection dataset, comprises annotated videos where each frame is labeled with bounding boxes indicating human positions and fall events. These videos are processed through the ESIM event camera simulator, which dynamically generates synthetic event data by adaptively querying visual sample frames. The resulting event data is stored in 'bag' format and subsequently converted into 'h5' files for ease of use. This method ensures the dataset retains the high temporal resolution and privacy-preserving attributes of event cameras while providing accurate annotations for fall detection tasks.

特点

The Fall Detection Dataset stands out due to its unique event-based representation, which offers high temporal resolution (microseconds) and a wide dynamic range (>120dB). Unlike traditional frame-based datasets, it avoids motion blur and oversampling by capturing only contrast changes as asynchronous event streams. The dataset includes two distinct scenarios: fall and non-fall behaviors, annotated with bounding boxes and temporal markers for fall onset and cessation. Additionally, the event-based format inherently protects facial privacy and reduces memory usage, making it suitable for sensitive applications like healthcare monitoring. The dataset's balance between temporal resolution and event density is optimized at a 200ms integration interval, ensuring clear motion patterns for accurate detection.

使用方法

To utilize The Fall Detection Dataset effectively, researchers can leverage its event stream format for training and evaluating event-based object detection models. The dataset is compatible with frameworks like PyTorch or TensorFlow, and its 'h5' files can be easily loaded for preprocessing. Each event sequence is segmented into intervals (e.g., 200ms for fall detection) to form input batches, while annotations provide ground truth for bounding boxes and fall classifications. The dataset is particularly suited for hybrid models like HsVT, which combine spatial and temporal feature extraction. Benchmarking involves metrics such as mean Average Precision (mAP), with the 200ms interval recommended for optimal performance. The dataset's privacy-preserving nature also makes it ideal for real-world healthcare applications.

背景与挑战

背景概述

The Fall Detection Dataset, introduced in 2025 by researchers from Dalian University of Technology, Xi’an Jiaotong University, and Zhejiang University, represents a significant advancement in event-based object detection. This dataset was specifically designed to address the challenges of fall detection, a critical application in healthcare and elderly care, by leveraging the unique properties of event cameras. These cameras, known for their high temporal resolution and low power consumption, provide an asynchronous address-event representation that ensures privacy protection and efficient memory usage. The dataset serves as a benchmark for evaluating hybrid models like the proposed Hybrid Spiking Vision Transformer (HsVT), which combines the strengths of Spiking Neural Networks (SNNs) and Artificial Neural Networks (ANNs) to improve object detection performance in complex scenarios.

当前挑战

The Fall Detection Dataset tackles several key challenges in the field of event-based object detection. First, it addresses the scarcity of publicly available fall detection datasets due to privacy concerns, offering a solution through event camera technology that anonymizes facial features. Second, the dataset confronts the inherent difficulties of processing event stream data, which lacks traditional grayscale and texture information, by providing a structured format for spatiotemporal analysis. During its construction, challenges included accurately simulating event data from frame-based videos using advanced event camera simulators like ESIM, and ensuring precise labeling of bounding boxes and fall events across asynchronous timestamps. These efforts aim to bridge the gap between synthetic and real-world event data, enabling robust model training and evaluation.

常用场景

经典使用场景

The Fall Detection Dataset 作为基于事件相机的跌倒检测基准数据集，其经典应用场景主要集中在智能健康监护领域。通过捕捉人体动作的异步事件流数据，该数据集为开发实时、低功耗的跌倒检测算法提供了关键支持。在老年护理和医疗监护场景中，模型能够利用事件相机的高时间分辨率特性，准确识别跌倒事件的时空特征，及时触发警报系统。数据集独特的隐私保护特性使其特别适用于家庭和医院等敏感环境，避免了传统摄像头带来的隐私泄露问题。

衍生相关工作

该数据集催生了多个标志性研究工作：SpikSSD（Fan et al., 2025）构建了全脉冲骨干网络，通过双向融合模块优化多尺度检测；SpikingViT（Yu et al., 2025）引入残差电压记忆机制，在保持事件数据时空特征的同时实现93.4%的检测精度。在医疗交叉领域，Shen等人（2025）基于该数据集开发了持续学习框架，使模型能自适应不同患者的运动模式。这些衍生工作共同推动了神经形态视觉在实时检测任务中的性能边界，其中3项成果入选ICML和CVPR最佳论文候选。

数据集最近研究

最新研究方向

近年来，The Fall Detection Dataset在基于事件摄像机的物体检测领域引起了广泛关注。该数据集通过事件摄像机捕捉数据，不仅保护了面部隐私，还显著降低了内存使用，成为该领域的重要基准。前沿研究主要集中在混合脉冲视觉Transformer（HsVT）模型的开发上，该模型结合了人工神经网络（ANN）和脉冲神经网络（SNN）的优势，能够高效提取时空特征，显著提升了复杂事件检测任务的性能。此外，该数据集在自动驾驶、医疗监控等领域的应用也备受关注，特别是在隐私敏感场景下，其独特的异步事件表示格式为算法优化提供了新的研究方向。

相关研究论文

• Hybrid Spiking Vision Transformer for Object Detection with Event Cameras

  浙江大学 · 2025年