Texture-less Object Detection And Recognition

Let computer have the object detection and recognition ability as people has always been the dream of human beings.Objects in real life can be divided into texture object and texture-less object,and texture object recognition technology has been mature,so our research is focus on texture-less object detection and recognition.However,object detection usually need to adapt to complex environment,such as clutter background,illumination change,scale and viewpoint change and partial occlusion.Researchers proposed many object detection algorithms trying to solve these problems.To realize a robust real-time detection of texture-less object,this paper mainly includes the following research work:Firstly,we extract object proposals to accelerate object detection and recognition.The tradition approaches utilized the sliding window paradigm to detect whether each window contains a target.Usually there will be a large number of backgrounds in the input image that do not contain the target,and the detection of the target in these areas will take a lot of time.This paper improves the existing object proposal extraction method of “Selective Search”,and use the improved algorithm to extract object proposals in the input image.Finally,the algorithm only identifies whether there is a target in these object proposals,which can greatly improve the object detection efficiency.Secondly,a method based on template matching is proposed.Texture-less object detection is usually based on matching the shape primitives extracted from object contour.The DOT method takes the dominant gradient orientation on texture-less object contour as features and then takes the sliding window approach matching template and input image.However,in complex backgrounds or there are interfering objects that are very similar to the target,the DOT method performance is not very good.This paper is extended on existing DOT methods.By setting a threshold for the matching similarity of DOT method,we can get all the detection windows whose matching similarity with template greater than the threshold.Then our algorithm extract local patch with rich distinction in the template and match the previous detected window again,which will exclude the wrong windows that do not really contain target objects.Thirdly,we implement a method for texture-less object detection and recognition through the convolution neural network.At present,the application of CNN is mainly concentrated on deep model.However,shallow network can be used when the target category is relatively small and the training data is limited.In this paper,two shallow CNNs are trained to detect texture-less object.The first network is used to classify different objects,the second network is used to map the object feature to low-dimensional space,and then use the nearest neighbor method in the low-dimensional space to find the template who has similar viewpoint and scale with the input image,So that we can capture both the object identity and 3D pose.Finally,in order to test the performance of methods in this paper,we have carried out many experiments in several databases with cluttered backgrounds,partial occlusion and objects with similar contour.The test results show that our methods can detect texture-less object in complex environment and have better performance in efficiency and robustness compared with several popular algorithms.