Year: 2023
Type: conference-paper
Abstract: In order to further improve the accuracy of object detection on datasets with time characteristics such as video and live streaming, this paper proposes an object detection algorithm that combines temporal neural networks. In this neural network, Mosaic data augmentation was used at the input end, and then entered the convolution section to obtain sufficient feature maps. These feature maps were continuously convolutionally processed to obtain feature maps of different scales for feature judgment. Features of different dimensions were normalized and sent to the temporal section. In the temporal part, the forgetting gate method was used to output the hidden layer information at the current time, which is a new feature map with preliminary judgment. Separate feature maps through 1 × 1 convolution expands the number of channels to obtain anchors with classification and detection information, and filters these anchors to screen positive and negative samples. After we generate prediction boxes for the target, we often generate a large number of redundant boundary boxes. Therefore, we need to remove boundary boxes with low positional accuracy and retain boundary boxes with high positional accuracy. Obtain the final prediction result. The model proposed in this article was compared with the YOLO and SSD models on the VisDrone dataset, and the improved algorithm achieved good accuracy improvement compared to the YOLO model and higher efficiency compared to the SSD model. The method proposed in this article provides a new approach and strategy for improving the integrity, accuracy, and universality of object detection tasks with temporal data.
Authors: Jiayin Teng, Guochao Fan, Yuejiao Zhang
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
OpenAlex ID: W4391316633