Cascading classifier with discriminative multi-features for a specific 3D object real-time detection

Abstract

Real-time specific 3D object detection plays an important role in intelligent service robots and intelligent surveillance fields. Compared to most existing approaches, which use simple template-matching methods, we present a novel discriminative learning-based method referred to as B-CST (BING - Colour + Shape + Texture) to detect a specific 3D object from a video in real time. Instead of the sliding-window technique, an original candidate extraction strategy is proposed, and that a new cascade classifier for recognition is also developed. In the candidate extraction stage, the rapid and high-quality objectness measure, binarised normed gradients, is modified to highlight the target candidate regions as well as to suppress undesirable background regions. In the recognition stage, each candidate region is then verified and further classified into different categories, which are denoted as positive, including multi-view images of target, or negative. The designed cascade classifiers conduct the recognition with discriminative multiple features, i.e. the novel dominant colour histogram, the histogram of oriented gradients and the original Gabor-CS-LTP feature, which is the centre-symmetric local ternary pattern of a special Gabor magnitude mapping. We evaluate our proposed method on our challenging new dataset consisting of 5 objects and two well-known public datasets and then compare it with other detection techniques for a single 3D object. A comparative study shows that our B-CST method is efficient in both high-quality detection results and detection speed, which can achieve the real-time processing requirements of video sequences (approximately 23 fps).