Multi-view Object Detection by Classifier Design and Interpolation
|School||Ocean University of China|
|Course||Applied Computer Technology|
|Keywords||Object Detection Multi-View Object Detection Model Design Model interpolation Boosting Strategy|
The detection of moving objects is a very active in recent years, image processing and computer vision field branch and topics at the forefront of concern. Has wide application value in the field of research, such as used in intelligent security monitoring and man-machine interface, human motion detailed analysis. The study on the basis of the summary and analysis of the domestic and international research work, for moving object detection in multi-view object detection, on the basis of the model design and model interpolation, to design a simple and effective multi-view object detection method, greatly improving the detection performance. First, manual intervention is introduced into the machine learning process, and is used to detect various perspectives on the model by manual design objects. In order to design a model more reasonable, designers in the design process can be based on the existing image and their understanding of the appearance of such objects to the model to adjust. This is a manual design classification process, the good to avoid the dependency of the classifier to the selected sample data. Given a plurality of perspective on the sample image, we chose several major perspective on the image as a training sample, manually designing a classifier for each viewing angle, respectively. Then, the designed classifier on several major perspective interpolation derived intermediate angles classifier. We assume that the base of each of the model feature weight and position can be regarded as the equation on the view angle, the interpolation process is the group in the model feature weights and position. Finally, we will design the classification obtained and interpolated derived classifier combined use Boosting strategy constructed object detecting a strong classifier for the multi-angle. Pedestrians and motor vehicles on the two data sets, respectively, after the multi-angle classifier classifier and the combination of the design and interpolation a single perspective detection experiments, and the results of their detection is obtained by the learning method corresponding classification the detection results were compared. The results show that the detection performance of the classifier design and interpolated derived are similar to the learning obtained in the detection performance of the classifier. The same time, the combined classifier design and interpolation derived from more than one perspective on the overall performance of the classifier greatly increased, which further validates the effectiveness of our design methods and interpolation methods.