Object Recognition on iCub Robot

摘要

Object recognition is a technology for finding and identifying objects in an image or a
\nvideo sequence. It relates to various fields including computer vision, machine learning, and
\nimage processing. Unlike other prior object recognition projects, this project mainly focuses on
\nproviding a visual recognition skill for iCub, a humanoid cognitive robot. A robot platform
\ncalled YARP (Yet Another Robot Platform) works as a middleware to support the
\ncommunication between hardware and software. This project uses supervised learning. Three
\ndifferent classes are trained and recognized by the robot: cup, sponge, and orange. For each
\nclass, 1800 sample images are collected from the cameras of the robot. Specifically, 80% of them
\nare used for training and validation, and the remaining 20% are used for testing. The training
\nmodel used for this project is neural network perceptron. Before feeding the features into the
\nmodel, preprocessing steps are implemented on those features; these steps include clipping the
\nimage to smaller size (160*120 pixels), converting the image to greyscale, and applying a
\nthreshold. In the training process, randomized weights and constant bias are used as parameters.
\nThe Sigmoid function, the Relu function and the Softmax function, as well as backward
\npropagation are used for training and updating class weights. The training results show that after
\n50 epochs the model converges to more than 90% accuracy. The testing results of the project
\nshow that the overall accuracy for the three classes is around 75% (results can be improved by
\nfollow-up modification). We can tell that the perceptron learning works well on this image set.
\nFurther modification can be done to increase the accuracy and the speed of training, such as
\nusing the stereo video to eliminate the noise and increasing the complexity of the neural network.
\nThis project can lead to more research such as feeding language signals to the robot or
\nimplementing reaching/picking after recognizing so that the robot will perform more like a human.