关键词: 人脸检测, 嵌入式系统, 深度学习, 果园

Abstract: In the process of orchard face detection, general face detection methods are difficult to accurately detect under the influence of complex backgrounds such as light spots and tree shadows, or cannot detect faces at a long distance when they have a certain detection accuracy. Deep learning methods can achieve highprecision, longdistance detection but require expensive equipment or realtime transmission of shooting information through a cloud platform for detection, which is too costly. In this paper, by constructing the MobileNet-YOLO lightweight deep learning face detection algorithm, it can improve the detection accuracy and detection distance while being suitable for embedded devices so that the device can directly perform face detection and reduce costs. After LBPH face recognition processing, the algorithm only saves and uploads pictures with detected faces to reduce storage and transmission pressure. The embeded device selected is Jetson Nano 2G, compared with ordinary embeded devices, GPU equipped on the basis of the same cost improves the computing power and is economical. The experimental results show that the detection distance of the face detection algorithm built in this paper is 8 meters. Compared with the SSD calculation, the detection speed is increased by 1.92 frames/s, the front face detection accuracy rate is 99.91%, and the side face detection accuracy rate is 96.66%. Compared with the YOLOv4-tiny algorithm, the front face detection accuracy rate is increased by 0.93%. The accuracy of side face detection is increased by 0.12%.

Key words: face detection, embedded system, deep learning, orchard