Soft-Weighted-Average Ensemble Vehicle Detection Method Based on Single-Stage and Two-Stage Deep Learning Models
Hai Wang, Yijie Yu, Yingfeng Cai, Xiaobo Chen, Long Chen, Yicheng Li
Abstract
The deep learning object detection algorithms have become one of the powerful tools for road vehicle detection in autonomous driving. However, the limitation of the number of high-quality labeled training samples makes the single-object detection algorithms unable to achieve satisfactory accuracy in road vehicle detection. In this paper, by comparing the pros and cons of various object detection algorithms, two different algorithms with a different emphasis are selected for a weighted ensemble. Besides, a new ensemble method named the Soft-Weighted-Average method is proposed. The proposed method is attenuated by the confidence, and it “punishes” the detection result of the corresponding relationship by the confidence attenuation, instead of by deleting the output of a certain model. The proposed method can further reduce the vehicle misdetection of the target detection algorithm, obtaining a better detection result. Lastly, the ensemble method can achieve an average accuracy of 94.75% for simple targets, which makes it the third-ranked method in the KITTI evaluation system.