Development of novel deep learning algorithms for object detection and recognition in humanoid robots

The humanoid robots are expected to facilitate people's lives in healthcare, houses and hotels providing various service support. Hence, it is important that the robots have object recognition capability. In that case, the humanoid robots should autonomously navigate and detect, recognize, and move all objects in the environment. However, object recognition is still a challenging problem at different locations and different object positions in real time. However, this is still difficult due to real-time control with different object positions and object views, and it is a problem to be solved. In this thesis, new algorithms are developed for object recognition and detection with humanoid robots. First, new classifier models with small structure based on Convolutional Neural Networks (CNNs) have been presented. Object recognition process was carried out with the proposed models. Second, new hybrid CNN-LSTM regressor models combining CNNs and Long Short-Term Memory Networks (LSTMs) have been proposed. Object movement application has been carried out with the Robotis-Op3 humanoid robot by using the method of learning from demonstration with the teleoperation method. Third, in the literature, a customized small new version of Pyramid Scene Parsing Network called PSPNet and a new version extracting the features using Discrete Wavelet Transform (DWT) has been proposed as DW-PSPNet. Moreover, DW-PSPNET is built from the encoder and decoder parts. Both parts include multiple originality. The encoder part proposes a novel Wavelet network applying DWT and including the spatial and channel attention blocks. In the decoder part, a new improved version of PSPNet is proposed. In experimental studies, firstly the object images have been taken with the camera of the humanoid robot. It has been shown that the proposed small structure classification models exhibit high accuracy recognition with lower number of parameters and shorter training time than complex models. Secondly, it has been observed that the proposed CNN models of hybrid CNN-LSTM based regressors show more accurate and stable results in the proposed system of humanoid robots learning to move objects from the demonstration. Third, the environment detection process has been performed by moving the humanoid robot in a house environment. It has been observed that the small structured PSPNet gives successful results in object segmentation with less parameter numbers than the conventional PSPNet. On the other hand, it has been shown that DW-PSPNet is faster and more successful than the conventional PSPNet and the popular networks in the semantic segmentation task in literature.