LightBlog
Contact at mumbai.academics@gmail.com or 8097636691/9323040215
Responsive Ads Here

Monday, 26 February 2018

Irregular Moving Object Detecting and Tracking Based on Color and Shape in Real-time System


Irregular Moving Object Detecting

 and Tracking Based on Color and 

Shape in Real-time System

Abstract—This paper describes an efficient approach for irregular moving object detecting and tracking in real-time system based on color and shape information of the target object from realistic environment. Firstly, the data is gotten from a realtime camera system at a stable frame rate. And then, each frame is processed by using proposed method to detect and track the target object immediately in consecutive frames. Finally, the target position based modifying controlling signal is used to control pan-tilt-zoom camera (PTZ camera) in order to automatically follow the target object. Our experiment results were obtained by using pan-tilt-zoom camera Sony EVI D70 under variety environments in real-time and our algorithm is verified that it can be implemented effectively and accurately at high frame speed, even 29.97 fps.
INTRODUCTION Real-time object detecting and tracking is an importance issue which aims to develop robots visual skills so that they are able to interact with a dynamic, realistic environment. The main challenges of the problem commonly are perspective, viewpoints changes, background clutter, image noise, scale, scene illumination and camera parameters. In the last few years, the problem has received a large amount of attention, in an attempt to improve the implementation at high frame rate with high accuracy. Color, gradient, intensity, depth were used to be effective features for object detection and recognition [1], contour and shape based approaches were also proposed. In the case of algorithm simplifier and reducing time consumption in order to be suitable with realistic environment, the two basic object features, color and contour information, should be taken for a job at hand. However, these two characters of object were used to be used separately for anti-jamming in weak systems, reduce consecutive images processing cost, improve working ability in complex environment, and etc. Elaborate contourbased methods were proposed, linking the edges, partitioning and connecting them to form a contour, then finding the sequence chains resembling the model outlines [2], learning detection from the segmented images, then applied for a larger un-segmented images set in [3], or using a bandwidth of a contour for deformable object [4]. They take, typically, at least a few seconds to scan and detect, therefore, they are far too expensive for real-time constraint system. Compared to object geometry, in most cases, color is a clearer identifying feature, less sensitive to noise and more largely robust to a view direction change and resolution. Hence, many color-based approaches were also developed. A simple and efficient performing algorithm, namely Backprojection, was introduced by Swain and Ballard [5], in which the pixels of the image are determined by their confidence values and the peaks in the confident space are considered as target objects. However, the applied area is the whole image so, if there are some regions in the background that has the same color as the target color, their confidence values are also high, but they are not the targets. This problem is solved in [6], in which higher weights are assigned to the pixels near the region center and lower to the background ones. These algorithms are simple, yet too computationally complex because of their complexity in putting effort into dealing with irregular moving target objects in a challenging environment. Our goal is to design an efficient, high accurate tracking and detecting system in which these above problems are mitigated, and it must run fast so that target object may be detected and tracked in real-time while consuming as few system resource as possible. Apparently, the most important thing here is high precise decision of target object and its localization, so, we have focused on both color and contour based detection and tracking. Nonetheless, the challenges of environment and time consumption should also be taken into account. Many above approaches were not appropriate to real-time system because of its complexity leading to reducing processing rate. To overcome this drawback, we divided the process into two main stages: detection stage and tracking stage. Let us call the detection stage the whole image processing and the tracking stage the interest region processing. Firstly, we use the whole image processing to detect object in the first consecutive frames. Until the system becomes stable, we use the obtained information of object position and size from the detection stage to decide the interest region which will be processed in the tracking stage in order to continue tracking our target. After a certain time, the process will be returned. The interest region processing is used to track the target while reducing process time consumption and the whole image processing is used to ensure the target object is tracked accurately even in the case the tracking stage cannot follow the object. Figure 1 summarizes our detecting and tracking approach in real-time system.

No comments:

Post a Comment