An Improved One-Stage Detector for Vehicle and Pedestrian Detection on Campus AGV

Ying Wang,Rui Li,Qi,Xiaojie Su
DOI: https://doi.org/10.1109/isr50024.2021.9419572
2021-01-01
Abstract:To solve the problem of introducing too many network parameters to a campus AGV (Automated Guided Vehicle) with limited computing resources, this paper proposes a modified deep convolutional neural network with the attention mechanism. The network is improved based on YOLOv4. Aiming to reduce the number of parameters and simplify the network structure for feature extraction, Dense-YOLOv4 is proposed, which effectively achieves the above goal while the detection accuracy is kept. In addition, we implemented an attention mechanism to the network to suppress features that are not particularly useful for the current task. The performance of the proposed network is verified through experiments on the Pascal VOC data and a customized dataset.
What problem does this paper attempt to address?