摘要：As an essential application in object detection, pedestrian detection has received extensive attention in many areas such as autonomous driving, video surveillance, and criminal investigation. With the rapid development of deep learning, pedestrian detection has made significant progress. When faced with multi-scale target pedestrians and dense crowds, false and missed detections are prone to occur, affecting accuracy. To overcome this problem, this study presents a multi-scale object detection network (Fast EfficientDet), based on an improved EfficientDet. Firstly, the backbone network EfficientNet is improved, and some of the deepwise separable convolutions that affect the speed of the model in the early training stage are discarded. At the same time, the Mish activation function is introduced to speed up the model's training. Secondly, a new feature pyramid-network Skip-BiFPN is proposed. Based on BiFPN, a cross-layer data stream is added to integrate the object's semantic and location information. In the face of complex environments, the network can better detect objects with large differences in size. Finally, the DIoU calculation method is introduced in the NMS post-processing. The suppression problem between the candidate frames is better handled by referring to the center point distance to solve object occlusion. Compared to the original EfficientDet series algorithm, the Fast EfficientDet-D0 obtained the best mAP of 84.98%, and the training speed increased by 15%. Compared to other algorithms, the Fast EfficientDet model has better performance.
