Humanoid robots find application in human-robot interaction tasks. However, despite their capabilities, their sequential computing system limits the execution of computationally expensive algorithms such as convolutional neural networks, which have demonstrated good performance in recognition tasks. As an alternative to sequential computing units, Field-Programmable Gate Arrays and Graphics Processing Units have a high degree of parallelism and low power consumption. This study aims to improve the visual perception of a humanoid robot called NAO using these embedded systems running a convolutional neural network. The methodology adopted here is based on image acquisition and transmission using simulation software: Webots and Choreographe. In each embedded system, an object recognition stage is performed using commercial convolutional neural network acceleration frameworks. Xilinx® Ultra96™, Intel® Cyclone® V-SoC and NVIDIA® Jetson™ TX2 cards were used, and Tinier-YOLO, AlexNet, Inception-V1 and Inception V3 transfer-learning networks were executed. Real-time metrics were obtained when Inception V1, Inception V3 transfer-learning and AlexNet were run on the Ultra96 and Jetson TX2 cards, with frame rates between 28 and 30 frames per second. The results demonstrated that the use of these embedded systems and convolutional neural networks can provide humanoid robots such as NAO with greater visual recognition in tasks that require high accuracy and autonomy.