Cited 0 time in webofscience Cited 0 time in scopus

Multi-class Vehicle Detection using Multi-scale Hard Negative Mining

Title
Multi-class Vehicle Detection using Multi-scale Hard Negative Mining
Authors
Kang, MinsungLim, Young Chul
DGIST Authors
Kang, Minsung; Lim, Young Chul
Issue Date
2019-11-20
Citation
The 6th International Conference on Internet of Vehicles (IOV2019), 109-116
Type
Conference
ISBN
9783030386504
ISSN
0302-9743
Abstract
The performance capabilities of object detection processes have been greatly improved due to the development of deep learning methods. As the performance of object detection methods improves, studies of problems that remained unsolved are now becoming more common. In CCTV technology, such as tracking technology, it has become easier to resolve the matching issue as the performance of object detection methods has improved. A network such as YOLOv3, a single stage multi scale based object detection method, robustly detects objects of various sizes while maintaining real-time performance. Object detection methods for multi scale structures are associated with the problem of an imbalance between a positive box and a negative box on each feature scale. In the CCTV environment, the object detection performance can be degraded due to this ‘unbalance’ problem because the number of objects corresponding to the positive box is relatively small. The learning time is also important because re-training is required for new environments that are constantly being added. In order to solve this problem, we propose a method that solves the unbalance problem through multi scale hard negative mining and that improves the object detection performance while also reducing the learning time.
URI
http://hdl.handle.net/20.500.11750/11435
DOI
10.1007/978-3-030-38651-1_11
Publisher
ASIA University, National Chung Cheng University
Related Researcher
Files:
There are no files associated with this item.
Collection:
Division of Automotive Technology2. Conference Papers


qrcode mendeley

Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.

BROWSE