An object detection approach with residual feature fusion and second‐order term attention mechanism

Automatically detecting and locating remote occlusion small objects from the images of complex traffic environments is a valuable challenging research. Since boundary box location not sufficiently accurate it difficult to distinguish overlapping occluded objects, authors propose network model with second-order term attention mechanism loss. First, backbone built on CSPDarkNet53. Then method designed for feature extraction based an item-wise mechanism, which uses filtered weighted vector replace original residual fusion adds reduce information loss in process accelerate convergence model. Finally, objected regression function studied problems missed detections caused by dense objects. Sufficient experimental results demonstrate that authors’ achieved state-of-the-art performance without reducing detection speed. The [email protected].5 85.8% Foggy_cityscapes dataset 97.8% KITTI dataset.