Bin picking robots require vision sensors capable of recognizing objects in the bin irrespective of the orientation and pose of the objects inside the bin. Bin picking systems are still a challenge to the robot vision research community due to the complexity of segmenting of occluded industrial objects as well as recognizing the segmented objects which have irregular shapes. In this paper a simple object recognition method is presented using singular value decomposition of the object image matrix and a functional link neural network for a bin picking vision system. The results of the functional link net are compared with that of a simple feed forward net. The network is trained using the error back propagation procedure. The proposed method is robust for recognition of objects.
A new algorithm for multi-object recognition and localization is introduced in this paper. This algorithm deals with objects which have different reflectivity factors and distinguish color with respect to the other objects. Two beacons scan multi-color objects using long distance IR sensors to estimate their absolute locations. These two beacon nodes are placed at two corners of the environment. The recognition of these objects is estimated by matching the locations of each object with respect to the two beacons. A look-up table contains the distances information about different color objects is used to convert the reading of the long distance IR sensor from voltage to distance units. The locations of invisible objects are computed by using absolute locations of invisible objects method. The performance of introduced algorithm is tested with several experimental scenarios that implemented on color objects.
In recent years, self-driving cars and reducing the number of accident casualties have drawn a lot of attention. Although it is crucial to increase driver awareness on the road, autonomous vehicles can emulate human driving and guarantee improved levels of road safety. Artificial intelligence (AI) technologies are often employed for this purpose. However, deep learning, a subset of AI, is prone to numerous errors, a wide range of threats, and needs to handle vast amounts of data, which imposes high-performance hardware requirements. This study suggests a deep learning model for object recognition that employs characteristics to describe data rather than images. Our model employs the COCO dataset as the training foundation, and it was suggested that the features be retrieved using the principal component analysis PCA extraction method. The current results demonstrate the efficacy and precision of our model, with an accuracy of 99.96 %. Furthermore, the performance indices, i.e., recall, precision, and F1-score, achieved about 1 for most of the COCO classes in training phase and promising results in testing phase.