中国科学院自动化研究所机构知识库

近些年科技工作者已经开发了多种电子辅助装置来帮助盲人及视障人士外出导航，但是现有的盲人导航装置存在价格昂贵、体验和交互性差等问题。针对盲人出行问题，伴随着大数据时代的到来和硬件计算能力的极大提升，本文主要采用计算机视觉及深度学习技术，研究面向盲人室外出行场景的物体检测算法并设计盲人视觉辅助眼镜系统，实现盲人出行的智能化和人性化。本文的主要工作和创新性成果包括：

1. 构建盲人室外场景数据集。本文沿盲道和交通路口采集相关图片数据集并进行标注，包括自行车、公交车、小汽车、摩托车、人、卡车和交通灯等七个常见物体类别，为后续的算法研究和盲人眼镜系统设计提供了数据基础。
2. 基于多尺度特征融合的物体检测算法研究。以提高物体检测算法的鲁棒性为目标，需要提取更为鲁棒的目标物体特征，并且从整幅图像中精确地检测多尺度目标物体。本文基于回归的深度学习物体检测算法，将深浅层的多尺度特征层融合以提取高语义和高分辨率特征相结合的复杂特征进行物体检测，并通过实验验证了本算法对物体检测的有效性。
3. 基于多层融合的小物体检测算法研究。为提高小目标物体的检测精度，需要从整幅图像中提取小物体更为丰富的几何细节信息。本文基于候选区域的深度学习物体检测算法设计多层融合的小物体检测模型，增加浅层卷积层用于物体检测，经过两次边框回归后精确地检测小物体，并在自建的数据集上验证了该算法的有效性。
4. 设计盲人视觉辅助眼镜系统。本文设计一款智能化和人性化的导盲眼镜来辅助盲人室外出行，一方面能够实现GPS定位和路线连续导航，另一方面通过双目摄像头采集前方场景信息，以物体检测算法识别障碍物和交通灯状态，并结合双目视觉定位算法进行避障，通过语音形式播报给盲人。

In recent years, scientists have developed a variety of electronic assisting devices to help blind and visually impaired people navigate outdoors, but current blind navigation devices are expensive, and experience and interaction are poor. Aiming at the blind travel problem, with the arrival of big data era and the great improvement of hardware computing ability, this paper mainly uses computer vision and deep learning technology to study the object detection algorithm for blind outdoor travel scenes and design blind vision-assisted glasses system, so as to realize the intelligence and humanity of blind travel. The main work and innovative achievements of this paper include:

1. Construct blind outdoor scene data set. This paper collects and labels related image data sets along blind roads and traffic intersections, including seven common object categories, such as bicycle, bus, car, motorcycle, person, truck and traffic light, which provides data foundation for subsequent algorithm research and blind glasses system design.
2. Research on object detection algorithm based on multi-scale feature fusion. In order to improve the robustness of the object detection algorithm, it’s necessary to extract more robust target object features and accurately detect multi-scale target objects from the entire image. In this paper, based on deep learning object detection algorithm of regression, the multi-scale feature layer of deep and shallow layer is fused to extract complex features combined with high-semantic and high-resolution features for object detection, and the effectiveness of the algorithm for object detection is verified by experiments.
3. Research on small object detection algorithm based on multi-layer fusion. In order to improve the detection accuracy of small target objects, we need to extract more detailed geometric details of small objects from the entire image. In this paper, we design a small object detection model of multi-layer fusion based on the deep learning object detection algorithm of candidate regions, and add shallow convolution layer for object detection, then detect small objects accurately after two bounding box regression. The effectiveness of the algorithm is verified on the self-built data set.
4. Design blind vision-assisted glasses system. This paper designs an intelligent and user-friendly guide glasses to assist blind travel outdoors. On the one hand, it can realize GPS positioning and route continuous navigation. On the other hand, it collects front scene information through binocular camera, then recognizes obstacles and traffic light status by object detection algorithm, and avoids obstacles combined with binocular vision position algortithm. Finally, it broadcasts to the blind by voice.