Research Article
An Object Detection and Tracking Algorithm Combined with Semantic Information
@INPROCEEDINGS{10.1007/978-3-030-89814-4_62, author={Qingbo Ji and Hang Liu and Changbo Hou and Qiang Zhang and Hongwei Mo}, title={An Object Detection and Tracking Algorithm Combined with Semantic Information}, proceedings={Mobile Multimedia Communications. 14th EAI International Conference, Mobimedia 2021, Virtual Event, July 23-25, 2021, Proceedings}, proceedings_a={MOBIMEDIA}, year={2021}, month={11}, keywords={Object detection and tracking Attribute recognition Similar object confusion Shared feature extraction module}, doi={10.1007/978-3-030-89814-4_62} }- Qingbo Ji
Hang Liu
Changbo Hou
Qiang Zhang
Hongwei Mo
Year: 2021
An Object Detection and Tracking Algorithm Combined with Semantic Information
MOBIMEDIA
Springer
DOI: 10.1007/978-3-030-89814-4_62
Abstract
This paper proposes a novel algorithm for object detection and tracking combined with attribute recognition that can be used in embedded systems. The algorithm, which is based on a single-shot multi-box detector (SSD) and kernelized correlation filter (KCF), can distinguish other objects similar to the tracking object, thereby solving the problem of confusion between similar objects. Different classification tasks in the multi-attribute recognition algorithm share the feature extraction module, which uses depthwise separable convolution and global pooling instead of standard convolution and fully connected (FC) layers, thereby improving the overall recognition accuracy and computational efficiency. Additionally, an attribute weight fine-tuning mechanism is added to improve the overall precision and ensure that different tasks are fully learned according to the degree of difficulty. Moreover, this algorithm reduces the size of the model without decreasing the accuracy, making it possible to be run on an embedded device. The results of experiments performed on OTB-100 demonstrate that a superior accuracy of 82.85% is achieved, and the precision and F1 indicator values reach 69.84% and 70.85%, respectively. The precision rate (PR) and success rate (SR) of the overall algorithm respectively reach 85.34% and 80.88%, which are higher than those achieved by the SSD algorithm. However, the size of the proposed attribute recognition algorithm is only 3.05 MB, which is about 1% of the size of other algorithms, indicating that the proposed algorithm not only improves the overall recognition accuracy, but also effectively reduces the model size.
