Vehicle Detection in Distorted Driving Video Based on Metric Learning and Single Shot MultiBox Detector

Abstract

With the gradually development of deep learning, the object detection algorithm has achieved remarkable applications, especially in the aspect of the automatic driving. Most of the object detection algorithms are used for pictures or videos obtained by a general camera. In practice, fisheye cameras are widely used, which will produce distorted image frames. The research of vehicle detection based on fisheye camera is relatively rare until now. If one network is trained on the existed public dataset, and tested on the distorted images or videos, the accuracy will decrease a lot. Thus, a distorted vehicle dataset needs to be manually labeled in the first. However, if we only use the distorted vehicle dataset to train the model, the mount of the distorted vehicle dataset is small, meanwhile the public datasets will not be fully used. On the other hand, the missing detection and false detection for the distorted images by using SSD algorithm is a considerable problem. Based on those considerations, firstly, transfer learning is adopted to transfer the parameters learned from the public vehicle dataset to the distorted vehicle dataset in this paper. Secondly, an algorithm named MLSSD for the distorted vehicle detection based on the labeled dataset is proposed to achieve a better performance for the vehicle detection, which mainly combines metric learning and SSD algorithm to enormously alleviate the missing detection and false detection. In addition, the scalable overlapping partition pooling (SOPP) method is proposed instead of the spatial pyramid pooling to achieve more robust feature map pooling. Experimental results show that the proposed MLSSD algorithm significantly outperforms other algorithms and achieves 88.3 % mAP on the distorted vehicle dataset, 3.1% more than the result obtained by the SSD network.