Occlusion-Aware Detection for Internet of Vehicles in Urban Traffic Sensing Systems
- PDF / 5,402,573 Bytes
- 7 Pages / 595.276 x 790.866 pts Page_size
- 13 Downloads / 226 Views
Occlusion-Aware Detection for Internet of Vehicles in Urban Traffic Sensing Systems Linkai Chen 1,2 & Yaduan Ruan 1 & Honghui Fan 2 & Hongjin Zhu 2 & Xiangjun Chen 2 & Qimei Chen 1 Accepted: 23 September 2020 # Springer Science+Business Media, LLC, part of Springer Nature 2020
Abstract Vehicle detection is a fundamental challenge in urban traffic surveillance video. Due to the powerful representation ability of convolution neural network (CNN), CNN-based detection approaches have achieve incredible success on generic object detection. However, they can’t deal well with vehicle occlusion in complex urban traffic scene. In this paper, we present a new occlusion-aware vehicle detection CNN framework, which is an effective and efficient framework for vehicle detection. First, we concatenate the low-level and high-level feature maps to capture more robust feature representation, then we fuse the local and global feature maps for handling vehicle occlusion, the context information is also been adopted in our framework. Extensive experiments demonstrate the competitive performance of our proposed framework. Our methods achieve better effect than primal Faster R-CNN in terms of accuracy on a new urban traffic surveillance dataset (UTSD) which contains a mass of occlusion vehicles and complex scenes. Keywords Vehicle detection . Vehicle occlusion . CNN . Context information
1 Introduction Vehicle detection from urban images or videos is a fundamental challenge in intelligent transportation system (ITS) [1–4]. The researches have studied in this field over several decades. The traditional methods such as Adaboost [5] and DPM [6] have got some achievements [7]. In recent years, CNN [8] contributed much to various computer vision problems include image classification [9], object detection [10, 11], semantic segmentation [12], etc. Due to the powerful representation of CNN, CNN-based detectors have got great progress. A lot of good algorithms have emerged, such as R-CNN [13], Faster R-CNN [14], SSD [15] and YOLO [16]etc. Vehicle detection from traffic surveillance videos is difficult due to the orientation, scale variation, occlusion, illumination change and complex scene [17, 18]. Figure 1 shows some examples for vehicle detection in our UTSD.
* Linkai Chen [email protected] 1
School of Electronic Science and Engineering, Nanjing University, Xianlin Road No.163, Nanjing 210023, China
2
School of Computer and Engineering, Jiangsu University of Technology, Zhongwu Road No.1801, Changzhou 213001, China
However, when applying CNN-based detectors to vehicle detection in urban video, one of the main challenges is that CNNbased detectors can’t deal well with occlusion and scale variation. The underlying reason of this problem is that CNN-based detectors only used global feature to detect and classify the vehicles. When occlusion occurs, the global feature involve background or the other category of vehicle. Therefore, the global feature can’t fully accurately represent the vehicle feature information. So we consider to fuse
Data Loading...
