Abstract: Motion object detection is an important research topic in the field of computer vision. Motion information is defined as the pixel position corresponding to the motion object in the image. However, in the context of autonomous driving, accurately extracting motion information is challenging due to changes in the image background caused by the motion of the onboard camera itself. Therefore, a motion information extraction model based on sparse optical flow estimation and deep learning was proposed to overcome the impact of background changes and detect motion information in the environment. The optical flow extraction module initially obtained global sparse optical flow by Shi-Tomasi corner detection and Lucas-Kanade (L-K) sparse optical flow estimation. The motion information discrimination module inferred suppression signals by inputting image depth information and sparse optical flow into a Transformer neural network, thereby suppressing the impact of background motion and extracting accurate motion information. The results show that the proposed method can extract motion information from images with an accuracy of 92%, which can be utilized for detecting moving targets for autonomous vehicles.

Key words: vehicle engineering; motion object detection; sparse optical flow; deep learning; autonomous driving

摘要： 运动目标检测是计算机视觉领域重要的研究内容,运动信息定义为运动目标对应图像中的像素点位置，然而在自动驾驶场景下，由于车载相机自身运动引起图像背景变化使得运动信息难以准确提取。提出了基于稀疏光流估计与深度学习的运动信息提取模型来克服背景变化带来的影响，检测环境中的运动信息。光流提取模块通过Shi-Tomasi角点检测及Lucas-Kanade（LK）稀疏光流估计初步得到全局稀疏光流；运动信息判别模块通过将图像深度信息和稀疏光流输入Transformer神经网络，推理出抑制信号，抑制背景运动带来的影响，从而提取出准确的运动信息。结果表明：该方法可以提取出图像中的运动信息，具有92%准确率, 可用于自动驾驶车辆检测运动目标。

关键词: 车辆工程；运动目标检测；稀疏光流；深度学习；自动驾驶