Adaptive Weighted Multi-Level Fusion of Multi-Scale Features: A New Approach to Pedestrian Detection

Great achievements have been made in pedestrian detection through deep learning. For detectors based on learning, making better use of features has become the key to their effect. While current efforts feature utilization improve performance, is still inadequate. To solve problem inadequate utilization, we proposed Multi-Level Feature Fusion Module (MFFM) and its Multi-Scale Unit (MFFU) sub-module, which connect maps same scale different scales by using horizontal vertical connections shortcut structures. All these are accompanied weights that can be learned; thus, they used as adaptive multi-level multi-scale fusion modules fuse best features. Then, built a complete detector, Adaptive Detector (AFFDet), an anchor-free one-stage detector make full for detection. As result, compared with other methods, our method performance challenging Caltech Pedestrian Detection Benchmark (Caltech) quite competitive speed. It state-of-the-art method.