PRED18 (PRED18: Predator/Prey DAVIS Dataset)

本文描述了一个应用
卷积神经网络 (CNN)
捕食者/猎物场景。 CNN 是根据来自的数据进行训练和运行的
动态和有源像素传感器 (DAVIS) 安装在
Summit XL 机器人（捕食者），它跟随另一个（捕食者）
猎物）。 CNN 由传统图像帧和
由常数组成的动态视觉传感器“帧”
DAVIS ON 和 OFF 事件。因此，网络是“数据驱动的”
以与场景活动成比例的采样率，因此有效
采样率从 15 Hz 到 240 Hz 不等，具体取决于机器人
速度。网络生成四个输出：向右、向左转向、
居中且不可见。在对标记数据进行离线训练后，
网络在运行的车载 Summit XL 机器人上导入
jAER 并实时接收转向方向。成功的
闭环试验结果，准确率高达 87% 或 92%
（取决于评估标准）被报告。虽然
提出的方法丢弃了精确的 DAVIS 事件时间，它
提供了与传统的兼容性的显着优势
深度学习技术，同时不放弃数据驱动计算的优势。

This paper describes a predator-prey scenario implemented with a Convolutional Neural Network (CNN). The CNN is trained and operated using data from a Dynamic and Active-pixel Vision Sensor (DAVIS) mounted on the Summit XL robot (the predator), which tracks another robot (the prey). The CNN takes as input both conventional image frames and dynamic vision sensor "frames" constructed from DAVIS ON and OFF events. As such, the network is "data-driven" with a sampling rate proportional to scene activity, so the effective sampling rate ranges from 15 Hz to 240 Hz depending on the robot's speed. The network outputs four discrete steering commands: turn right, turn left, center, and invisible. After offline training on labeled datasets, the network is deployed on the onboard system of the active Summit XL robot, leveraging the jAER toolkit to generate steering directions in real time. Successful closed-loop experimental results are reported, with accuracy reaching up to 87% or 92% depending on the evaluation metric. While the proposed method discards the precise timing of DAVIS events, it offers significant advantages in compatibility with conventional deep learning techniques without sacrificing the benefits of data-driven computation.

搜集汇总

背景与挑战

背景概述

PRED18是一个用于捕食者/猎物场景的目标跟踪数据集，基于卷积神经网络和动态视觉传感器技术，在机器人上实现实时转向预测，准确率最高可达92%。该数据集以数据驱动方式运行，采样率随场景活动动态调整，适用于CV目标跟踪任务。