Indoor and Outdoor Face Recognition for Social Robot, Sanbot Robot as Case Study

• Published in: 2020 28th Iranian Conference on Electrical Engineering (ICEE) pp. 1 - 7
• Main Authors: Ashtari, Erfan, Basiri, Mohammad Amin, Nejati, Saeid Mohammadi, Zandi, Hemen, Rezaei, Seyyed Hossein SeyyedAghaei, Masouleh, Mehdi Tale, Kalhor, Ahmad
• Format: Conference Proceeding
• Language: English
• Published: IEEE 04.08.2020
• Subjects: Cameras; Deep Neural Network; Face recognition; Human-robot interaction; Robots; Sanbot robot; Servers; Siamese Neural Network; Social robot; Streaming media; Testing; Training
• ISSN: 2642-9527
• DOI: 10.1109/ICEE50131.2020.9260698

The interaction between human and robots is of paramount importance in comforting robot and human in the context of social demand. For the purpose of human-robot interaction, the robot should have the ability to perform a variety of actions including face recognition, path planning, etc. In this paper, face recognition has been implemented on the Sanbot robot. Since the Sanbot robot is intended to work in real environment, therefore indoor and outdoor environment is taken into account in proposing the corresponding face recognition algorithm. For each case a robust pre-processing algorithm should be designed and which can circumvent a challenging problem in face recognition, namely, different lighting conditions (light intensity, angle of radiation, etc.). In case of indoor environment, faces in an captured image by the robot HD camera are found using a Haar-cascade algorithm. Afterwards, a histogram equalization is applied to face images in order to standardize them. Then commonly practiced Deep convolutional neural network structures such as Inception and ResNet are used to design a model and trained end-to-end on a customized dataset with strong augmentation. Finally, by using a voting method, proper prediction is carried out on each face. In what concerns the outdoor environment, which has more challenges, upon applying histogram Equalization on the captured image, faces are found using a MultiTask Cascaded Convolutional Neural Network. Then face images are aligned as head orientation are corrected. Finally, cropped face image is fed to Siamese Network in order to extract face features and verifying individuals. From several practical results it has been inferred that the accuracy of the indoor method is nearly 93% without voting and with voting 97%, and the outdoor method is about 95%.