智能制造技术与系统研究中心知识库

随着人工智能技术的飞速发展，各个领域都迎来了“无人化”时代。在智能交通领域中，无人车系统技术正在逐步渗入到国防建设、工业生产和日常生活中。在无人车系统中，行人和车辆检测是无人车系统运行的基石。目前，无人车系统中主要采用以激光雷达技术为主、其他技术为辅，对车辆和行人等进行检测。但是，激光雷达技术存在着成本较高和粗糙物体难以分类等缺点。

与传统物体检测算法相比，基于深度学习的物体检测算法具有更高的精确度和更强的广泛性。不过，它检测精度的提高往往导致检测速度的下降。同时，它对支撑硬件的要求造成了其在一定程度上具有应用局限性。在实际交通场景中，行人和车辆的检测任务具有以下四个特点：1）交通场景复杂多样；2）物体之间易出现遮挡或形变；3）远处物体不易被检测；4）检测需要高精度与高效率。

针对基于深度学习的物体检测算法的优势和局限性，考虑实际场景中对行人和车辆检测的特点，本文预优化现有的通用物体检测的算法Faster-RCNN，展开面向低速无人车的行人和车辆检测算法研究。本文的主要工作和主要内容总结如下：

1）提出了一种基于深度卷积神经网络的特征提取算法。该算法通过对空间和局部信息的提取以及对不同分辨率特征层的融合，使得输入图像通过特征提取网络得到的特征，具有更多细节信息。将该优化算法应用在resnet18 和 vgg16的基础网络上，在数据集Pascal_voc上进行实验，表明了优化算法的有效性。

2）提出了一种生成候选区域框并增强的算法。该算法利用anchor string机制代替传统候选区域生成机制，提高了尺寸较小物体的检测精度；同时，基于anchor string机制，提出了anchor-context机制，根据生成候选区域框的高、宽、面积等信息，自适应地增加语义信息，从而提高遮挡物体的检测精度。将优化算法在数据集Pascal_voc上进行实验，表明了优化算法的有效性。

3）提出了一种基于深度卷积神经网络的行人和车辆检测算法。我们在通用特征提取网络的基础上，合理融合上述的优化方法，提出了一种基于深度卷积神经网络的行人和车辆检测算法。这种算法可以提升对小物体与被遮挡物体的检测精度；在速度上也可以满足低速无人车，比如无人巡逻车的实时检测的速度要求，模型的大小满足实际模型移植的要求。将优化算法在数据集上进行实验，表明了该优化算法的有效性。将优化算法移植到无人巡逻车上进行实景实验，验证了算法在实际场景中的有效性。

With the rapid development of artificial intelligence technology, all fields have ushered in the era of "artificial intelligence".In the field of intelligent transportation, unmanned vehicle system technology is gradually infiltrating into national defense construction, industrial production and daily life. In the unmanned vehicle system, vehicle and pedestrian detection is the cornerstone of the unmanned vehicle system. At present, lidar technology is mainly used in the unmanned vehicle system, and other technologies are supplemented to detect vehicles and pedestrians. However, lidar technology has some shortcomings, such as high cost and difficulty in classifying rough objects.At the same time, with the development of deep learning, object detection algorithm based on deep learning has gradually matured and improved.

Compared with traditional object detection algorithms, object detection algorithm based on deep learning has higher accuracy and more extensive. However, the improvement of its detection accuracy often leads to the decrease of detection speed. At the same time, its requirement for supporting hardware has resulted in its application limitation to a certain extent. The detection tasks of pedestrians and vehicles in actual traffic scenarios have the following four characteristics: 1) traffic scenarios are complex and diverse; 2) occlusion or deformation between objects is easy to occur; 3) there are objects that are difficult to detect in the distance; 4) real-time performance is needed while ensuring accuracy.

Considering the advantages and limitations of object detection algorithm based on deep learning and the characteristics of pedestrian and vehicle detection in real scenes, this paper is going to optimize the existing general algorithm Faster-RCNN based on deep convolution neural networks and develops pedestrian and vehicle detection algorithm for unmanned vehicles with low speed in the park. The main work and contents of this paper are summarized as follows:

1) Propose a feature extraction algorithm based on deep convolution neural networks. By extracting spatial and local information and fusing different resolution feature layers, the algorithm enables the input image to get more detailed information through feature extraction network, and better complete the detection task of pedestrians and vehicles. The optimization algorithm is applied to the basic network of resnet18 and vgg16. Experiments show the effectiveness of the optimization algorithms on the datasets.

2) Propose an algorithm for generating and enhancing anchor box. In this algorithm, anchor string mechanism is used to replace the traditional anchor box generation mechanism, which increases the detection accuracy of smaller objects with less complexity. At the same time, anchor-context mechanism is proposed by using anchor string mechanism. In other words, we use the information of height, width and area of candidate anchor box to generate self-adaptively context information. In order to improve the detection accuracy of occluded objects. Experiments show the effectiveness of the optimization algorithms on the datasets.

3) Propose a pedestrian and vehicle detection algorithm based on deep convolution neural network. Based on the general feature extraction network and the above optimization methods, we propose a pedestrian and vehicle detection algorithm based on deep convolution neural network. This algorithm can improve the detection accuracy of small objects and occluded objects; it can also meet the speed requirements of real-time detection of slow-speed unmanned vehicles, such as unmanned patrol vehicles; at the same time, the size of the model can meet the actual model transplantation. Experiments on data sets demonstrate the effectiveness of the optimization algorithm. At the same time, the optimization algorithm is transplanted to the unmanned patrol vehicle to carry out field experiments, which verifies the effectiveness of the algorithm in the actual scene.