Urban Object Detection

近年来，我们看到使用基于深度学习的对象检测器的应用程序数量大幅增长。自动驾驶辅助系统 (ADAS) 是影响最大的领域之一。这项工作提出了一项新颖的研究，评估了用于城市物体检测和定位的最新技术。特别是，我们调查了更快的r-cnn方法在各种户外城市视频中检测和定位城市物体的性能，这些视频涉及在场景中移动的行人、汽车、自行车和其他物体 (城市驾驶)。我们提出了一个新的数据集，用于对实时对象检测器 (更快的r-cnn) 的准确性进行基准测试。部分数据是使用安装在车辆上的高清摄像机收集的。此外，某些数据的注释较弱，因此可以用于测试弱监督学习技术。已经存在城市对象数据集，但是它们都不包括所有基本的城市对象。我们进行了广泛的实验，证明了基线方法的有效性。此外，我们提出了一种r-cnn plus跟踪技术，以加速实时城市物体检测的过程。

In recent years, we have witnessed substantial growth in the number of applications utilizing deep learning-based object detectors. Advanced Driver Assistance Systems (ADAS) is one of the most impactful domains. This work presents a novel study evaluating state-of-the-art technologies for urban object detection and localization. Specifically, we investigate the performance of the Faster R-CNN method for detecting and localizing urban objects across various outdoor urban videos, which feature pedestrians, cars, bicycles, and other moving objects in the scene (urban driving). We propose a novel dataset for benchmarking the accuracy of real-time object detectors, specifically the Faster R-CNN. Portions of the data were collected using high-definition cameras mounted on vehicles. Additionally, some of the data is weakly annotated, making it suitable for testing weakly supervised learning techniques. While existing urban object datasets have been developed, none of them encompass all fundamental urban objects. We conducted extensive experiments that validate the effectiveness of the baseline method. Furthermore, we propose an R-CNN-plus tracking technique to accelerate the real-time urban object detection pipeline.

搜集汇总

背景与挑战

背景概述

该数据集是一个用于自动驾驶城市物体检测研究的数据集，包含车辆采集的城市场景视频数据，支持弱监督学习技术测试，并提出了加速实时检测的R-CNN加跟踪方法。