In Mathematical biosciences and engineering : MBE
Nowadays, object detection methods based on deep neural networks have been widely applied in autonomous driving and intelligent robot systems. However, weakly perceived objects with a small size in the complex scenes own too few features to be detected, resulting in the decrease of the detection accuracy. To improve the performance of the detection model in complex scenes, the detector of an improved CenterNet was developed via this work to enhance the feature representation of weakly perceived objects. Specifically, we replace the ResNet50 with ResNext50 as the backbone network to enhance the ability of feature extraction of the model. Then, we append the lateral connection structure and the dilated convolution to improve the feature enhancement layer of the CenterNet, leading to enriched features and enlarged receptive fields for the weakly sensed objects. Finally, we apply the attention mechanism in the detection head of the network to enhance the key information of the weakly perceived objects. To demonstrate the effectiveness, we evaluate the proposed model on the KITTI dataset and COCO dataset. Compared with the original model, the average precision of multiple categories of the improved CenterNet for the vehicles and pedestrians in the KITTI dataset increased by 5.37%, whereas the average precision of weakly perceived pedestrians increased by 9.30%. Moreover, the average precision of small objects (AP_S) of the weakly perceived small objects in the COCO dataset increase 7.4%. Experiments show that the improved CenterNet can significantly improve the average detection precision for weakly perceived objects.
Zhou Jing, Chen Ze, Huang Xinhan
2022-Sep-01
** CenterNet , anchor-free , attention mechanism , multi-scale feature enhancement , object detection **
