Detection proposal method based on shallow feature constraints

Rapid detection of small or non-salient attacking objects constitutes the dominant technical concern for prevention of airport bird strike. According to changes of the object observed from far to near, a novel detection proposal method based on shallow feature constraints (ShallowF) is thus proposed. Specifically, the object is located approximately by virtue of feature points, narrowing search spaces, reducing the number of sampling frames, and improving the efficiency of detection proposals. Then sampling rules are specified by connected domains and feature points, further narrowing search spaces and reducing the number of sampling frames. Finally, based on the difference between the target contour and the background, the structured edge group in the bounding boxes is extracted as the scoring basis for target detection before test and validation on the COCO Bird Dataset [1] and the VOC2007 Dataset [2]. Compared with the most advanced detection proposal methods, this method can improve the accuracy of candidate bounding boxes while reducing their quantity.