A novel monocular object detection and localization framework based on inverted multi-scale attention

Abstract The observation angle of camera to target is non-orthogonal in wide applications of computer vision, which could result in the deviation between the image coordinates of the target and its actual position in the real world. This paper presents a novel monocular visual object detection and localization algorithm to address such a challenge to improve detection accuracy and reduce projection distortion in non-orthogonal camera angles. First, an inverted Efficient Multi-scale Attention (iEMA) module is proposed to strengthen the recognition capability of the algorithm for small objects. Second, a new feature extraction module, Dynamic Head Plus (DBS+), is created to better handle targets of different shapes. Third, Soft Intersection over Union (SIoU) is adopted as the bounding box regression function, enhancing the convergence speed and detection accuracy. Fourth, a nonlinear localization model is constructed to achieve monocular visual target localization by introducing angular increment. Finally, experiments are conducted to verify the proposed algorithm. Results show that the improved algorithm can enhance detection accuracy by 8.6%, recall by 5.3%, mAP by 13.3%, reduce the edge error by 3.4, and decrease the localization error by 9%. This paper is of great significance for improving the performance of monocular vision in non-orthogonal observation scenarios, and provides more effective algorithm support for the field of mobile object detection and localization.