We found a match
Your institution may have rights to this item. Sign in to continue.
- Title
A lightweight network for vehicle detection based on embedded system.
- Authors
Wu, Huanhuan; Hua, Yuantao; Zou, Hua; Ke, Gang
- Abstract
The target detection algorithm has a high accuracy rate for the detection of general objects. However, for specific kinds of objects, especially those with complex application scenarios and high requirements on recognition rates, the accuracy and real-time performance are still far behind the actual application requirements. Meanwhile, the high real-time performance of most current target detection algorithm depends on high-performance computers or Graphics Processing Units (GPUs), which limits their application in low-performance embedded system. In this paper, an embedded vehicle detection algorithm was proposed. Compared to the existing general detection algorithms, the network structure of the proposed algorithm has been improved in many aspects. Firstly, the overall network structure of YOLOv3-tiny has been improved, and the way of feature extraction has been modified to increase the network in terms of detection speed and recognition accuracy. Secondly, a spatial pyramid pooling structure is introduced into the network to enhance the learning capability, and the spatial pyramid pooling structure is also improved by increasing the number of pooling layers in the structure and cascading multiple groups of pooling layers of different sizes. Thirdly, to speed up target detection process, the original clustering algorithm is replaced by Kmean++ clustering algorithm, and the original IoU (The Intersection over Union) loss function is replaced by GIoU (Generalized IoU). Fourthly, according to the scale factor of batch normalization layer, the model is compressed and trimmed to remove redundant channels, and a small model more suitable for embedded system application is obtained. Finally, Raspberry Pi 4B and Inter NCS2 neural computing stick are used as the embedded system with the input image size of 416 × 416. Experimental tests on UA-DETRAC dataset obtain higher mAP than other advanced existing methods, indicating that the network model proposed in this paper could effectively improve the real time and accuracy of multi-vehicle target detection.
- Subjects
PYRAMIDS; GRAPHICS processing units; OBJECT recognition (Computer vision); FEATURE extraction; RASPBERRY Pi; COMPUTER graphics
- Publication
Journal of Supercomputing, 2022, Vol 78, Issue 16, p18209
- ISSN
0920-8542
- Publication type
Article
- DOI
10.1007/s11227-022-04596-z