| Summary: | Object detection and distance estimation based on videos are important issues in advanced driver-sssistant system (ADAS). In practice, fisheye cameras are widely used to capture images with a large field of view, which will produce distorted image frames. But most of the object detection algorithms were designed for the nonfisheye camera videos without distortion, which is not suitable for the application of ADAS since one always expects the panorama stitching and object detection system should share one set of cameras. The research of vehicle detection based on fisheye cameras is relatively rare. In this paper, vehicle detection and distance estimation based on fisheye cameras are studied. First, a multi-scale partition preprocessing is proposed, which can enlarge the size of small targets to improve the detection accuracy of small targets. Second, parameters learned from the public datasets without distortion is transferred to our fisheye video dataset. Then metric learning-based single shot multibox detector (MLSSD) is proposed to improve the accuracy of distorted vehicle detection. Combining metric learning and SSD network, MLSSD can significantly reduce the missing and false detection rates. Moreover, a scalable overlapping partition pooling method is proposed to explore the relations among the adjacent features in a feature map. Finally, the distance between the driving vehicle and vehicles around this vehicle is estimated based on the object detection results by the method of marker points. Experimental results show that our proposed MLSSD network significantly outperforms other networks for distorted object detection.
|
|---|
