Object Detection Based On Deformable Part Model With Context

Object detection and recognition is a very active research direction in the field ofmachine learning, pattern recognition and computer vision. It has been widely used inmany areas, including face recognition,pedestrian detection, intelligent video analysis,pedestrian tracking, in security and protection system, including traffic scene objectrecognition, vehicle counting, retrograde detection, license plate detection andrecognition in traffic areas, including content based image retrieval in the field ofInternet, including target object recognition in the field of the robot and so on. Objectdetection has been applied to every aspect of human life, have changed and willcontinue to change the life style of human society.Deformable part-based model is a remarkable algorithm in object detection.Since released, it has received extent attention from reseachers. Many improvementswere proposed. After50years, remarkable development has been achieved in objectdetection and recognition area. Since2007, DPM and many involved algorithms haveachieved excellent performance in the task of PASCAL VOC object recognition anddetection.Some limitations exist in the detection algorithm based on appearance. Thequality of the image, as well as the changes of object stucture, will influence the resultof detection and recognition.Reseachers found that certain object was releavent tocertain scene, and the utilization of scene information (context information) iscontribution to improve the accuracy of object detection and recognition.The goal of our thesis is to utilize the scene context to improve the accuracy ofDeformable part-based model.First of all, the gist algorithm is the best choice to extract scene information. Thisalgorithm utilizes Gabor filters to process image to get oritation, color, and intensityinformation. And then the color and intensity information are applied toCenter-Surround algorithm. At last, information from these channels is integrated intoa feature vector called gist. This is an excellent algoritm to discribe the color, intensity and texture information for image. This algorithm is applied to extract scene contextfrom image.Then we try to use Feature context algorithm to extract scene information. Codewords were extracted by Densesift algorithm, and then used to code every pointfeature in image under the principal of Radial Basic Coding. All the point features aredivided into grids of circle with the reference point as the centre. The advantage ofthis algorithm is that the spatial information is well reserved. This algorithm is alwaysused to extract local context of object, but in this article is used to extract globalcontext as a try.The last research goal is to integrate the context information and DPM. Thedimension of the global context feature is very high, which will bring more noise. Soit will be reduced with the help of Principal Component Analysis algorithm. Theoutput of DPM in training set should be normalized including the normalization ofcoordinates of the bounding box and the score. Then, these two parts of informationfrom different algorithm are integrated to a feature vector as a whole.At the training stage, DPM algorithm and gist algorithm run independently intraining dataset. These two parts are integrated into a new feature vector. And then,the training samples will be carefully choosed under a threshold to form a newtraining dataset, which trained by Support Vector Machine to get a new classifier.At the testing stage, the new classifier is used to revise the output of DPM invalidation dataset.At the experiment stage, three catogories are detected in PASCAL VOCdataset. And then the outputs are evaluated by the Average Precision (AP). Theresults prove that our new algorithm work well.