Non-Rigid Object Detection Via Fast One-Class Model

Abstract

One-class classification (OCC) has been successfully used in various applications. However, limit to the non-robustness, non-sparsity, or high-order optimization, existing methods fail to detect non-rigid objects, especially for a high-precision real-time detection task in precision forestry and agriculture. In this study, we propose a fast one-class model using L1 norm metric. Computationally, the proposed model is led to a first-order optimization. In order to address the issue of nonlinearization, we provide an alternative optimization strategy. Owing to directly minimizing ||α||_1, the resulting solution α is inclined to be highly sparse, which is particularly advantageous for real-time detection. Furthermore, we are also introducing two additional acceleration algorithms to further improve training speed. Finally, extensive experiments for non-rigid interest object detection are carried on public and our collected images and videos. Compared with SOTA (state-of-the-art), our proposed method demonstrates the superiority in terms of non-rigid object detection accuracy rate and error warning rate, as well as in training time and real-time ability.

Keywords: One-class classification, L1 norm, real-time, high-precision, Object Detection