A Method For Object Recognition From Multi-VIEWS For Mobile Devices

Object recognition has always being the most important research field in computer vision. Mobile networks are stepping from3G to the era of4G, coupled with the maturity of mobile internet industry, which enables faster data transition on mobile devices. Besides, technical advancement in chip making and smart devices making bring us excellent handheld devices with faster CPU, more RAM, great dpi(dots per inch) screen and high resolution camera. This makes a lot of systems which could not be executed on mobile devices now could be used on smart phones, tablets, etc. Those devices just tipped the tipping point of massive needs of various scenarios, which inspires abundant mobile applications, of course, there are many involves image processing and object recognition, and those applications would not be only run on smart phones and tablets.Year2013could be named as the first year of wearable devices, one big class of wearable devices is AR(Augmented Reality) devices. AR systems need to capture the scene through camera, then recognize this picture with object detection and object recognition, then track the detected targets, multi-media information(like wiki content or google search results about the object) will then be displayed besides/on this target on screen virtually. This paper proposed an approach for multi-view object recognition could be deployed on mobile devices. To recognize object in images taken from multiple views, the template collects5standard image from5fixed views (front view, back view,2side views and aerial view), then match the input image with the template serially, an object recognized once one template image matches the input successfully with our new approach, which could increase the rate of identification.There are different interferences in real scenes, eg. complicated background or main part of the image is not the target object, which mischiefs the recognition results. We generate several candidates areas by introducing new object detection method Objectness BING and combining analysis on color contrast. Recognition is then run on those promising candidates, which saves us computation work and improves recognition results。A very important measurement of this method is speed, for we want this method could be deployed on mobile devices in the future. We choose ORB local binary feature methods for its good enough performance and excellent speed after study and experiment on popular local feature algorithms. We use color histogram in HSV space to enhance the local feature match result and to recognize object jointly.This paper design, and test the recognizing the method on a laptop, and test simple deployment on Android device. We achieve very good recognizing ability for objects in template, which provides a new solution for object recognition.