Object Detection And Grasp Based On Computer Vision

Object detection and grasp has extensive applications in industries and people's daily life,such as intelligent transportation,video surveillance,industrial automation,and so on.Object detection has been studied for many years and many methods were proposed.However,for multi-modal object in complex background,the detection accuracy has not achieved the desired level.Object grasping has been widely used in industry,but most of these applications tend to grasp very simple objects.It is difficult to grasp complex objects for these methods.This thesis proposes a new method based on multiple visual features and extreme learning machine(ELM)algorithm.This proposed method mainly includes two parts:Multi-modal object detection and grasp detection.For the multi-modal object detection,an object-background classification system is proposed to detect the objects in the image based on sliding window method.Histograms of oriented gradient(HOG)features and histograms of color features are extracted as the feature representations during the training and detecting processes.The first contribution of this object detection system is that a two-stage serially cascaded detection model is built to detect the multi-modal objects.The second contribution of this object detection system is that the classifiers are trained by using ELM algorithm.ELM algorithm can deal with non-linear problem which is caused by the complex background.For the grasp detection,a rotary sliding window detection framework is proposed to detect the grasp of an object.ELM kernel algorithm is used to train classifiers and HOG features are extracted as feature representations.The first contribution of this grasp detection system is that an object's major orientations algorithm is proposed to improve the efficiency of the detection system.The second contribution of this grasp detection system is that a two-stage detection model is constructed by ELM kernel algorithm to detect the complex object's grasp.To evaluate this proposed object detection method,the experiments are taken on the PASCAL Dataset.In this dataset,the objects and the background are very complex.This can extremely enhance the difficulty of the object detection.The detection result shows that this proposed object detection system achieves high precision on that dataset.For the grasp detection system,Cornell grasp detection dataset is used to evaluate this proposed method.This dataset has a lot of complex objects including cups,glasses,screwdrivers etc.The detection accuracy of this proposed method can achieve about 64.8%.This accuracy is better than some algorithms using deep learning.This proposed object detection and grasp detection system is also evaluated on the manipulator platform.The experimental results show that this proposed method is able to successfully execute a grasp in 86%of the trails.