Indoor Object Detection With Convolution Neural Networks For Mobile Robot

Visually recognizing objects is one of the fundamental ability to have if robots need to work along with humans and in the environment designed for humans.In past computer vision systems although performed well in controlled environment such as industry settings were not suitable for diverse use and required considerable human efforts in training.Recent advances in machine learning techniques and its application in computer vision has resulted in unprecedented improvements in real-world natural image classification but its reliance on large training data(sometimes millions of training images)with long training time(hours to days)make these machine learning algorithms less suitable for real-time general purpose robotics.In this study,we have used Convolution Neural Network(CNN)for object recognition task(given image,recognize the object present in image).To avoid the long and complicated training procedure,we feed captured image into pre trained CNN and use output of CNN as feature vector that represent the object in the image.As the object is being tracked by camera,sequence of feature vectors is stored in memory and eventually Support Vector Machine(SVM)is trained to classify the objects later on.SVM can be trained in reasonably small time with small training data hence we can achieve real-time performance.Using the technique mentioned above we investigate the accuracy of such vision system in real world indoor environment.Using camera and motion detection,our system tracks the object presented in field of view.If learning mode is activated by user,system trains a new SVM for this new object being tracked.If prediction mode is activated by user,system try to guess the identity of the object based on already trained models.In another scenario we investigated the possibility of classifying objects using only visual feedback.System checks if the feature vector of the given object is similar to any known object feature vector by measuring Euclidian distance between feature vectors.Finally we have compared our approach with other similar object recognition algorithms using standardized dataset.