Research On Object Detection Technology Based On Deformable Part Model

Vision-based object detection technology can be directly applied to many fields, such as visual surveillance, automotive safety, human machine interaction and interactive entertainment. It has broad commercial application value and market prospects, and thus becomes hot research in the field of computer vision and image processing. In recent years, with the development of science and technology, the object detection technology has achieved significant progress. In more complex scenes, its performance is still not satisfactory, large object changes, light,occlusion and other effects render object detection a still challenging task.The challenge is mainly presented as: the detection accuracy of the current algorithm is still low, its detection rate is still slow, which means there is still a wide gap from practical application. In addition, how to balance accuracy and real-time in object detection is still a challenging research topic.Deformable Part Model (DPM, for short) has recently received great attention in the field of generic object detection. This detection algorithm represents the object by a set of local parts and the positional relationship between each part, thus has a good effect in dealing with object occlusion and intra-class variation. In the study and research of DPM object detection algorithm, it is found that the DPM object detection algorithm has higher detection precision at the cost of large computation, which limits its application development. This paper mainly focuses on the improvement of DPM detection speed, proposes the algorithm of coarse-to-fine detection and the fast build of characteristic pyramid, and on the basis of this, the characteristic pyramid is layered detection. The main work and innovation of this paper are as follows:Firstly, we explain the theory and implementation details of DPM-based object detection in detail.Secondly, we combine interest extraction with DPM target detection algorithm to form a coarse-to-fine object detection method, the detection speed can be greatly improved by using the method to preprocess the image and obtain the recommended window for DPM.Thirdly, aiming at the problem of single HOG feature, we propose a multi-feature complementary fusion method to detect the object and improve the accuracy of the target detection. At the same time,approximate calculation method is used to construct the characteristic pyramid, which greatly reduces the computational complexity burden.Lastly, in the aspect of filter response, we propose a hierarchical detection algorithm. By using the characteristics of the root model, the feature layer with potential targets is found quickly, and then the part model is used to detect whether there are targets. The computational burden associated with the use of root model and part models for each layer is avoided.In summary, this paper conducts a study on the DPM-based object detection task, and then proposes a corresponding optimization algorithm for the problems of slow detection speed and low detection precision.Experimental results show that this method can effectively improve the effect of DPM-based object detection and improve the detection speed significantly while maintaining good detection accuracy.