The application of object detection technology has been used in people's daily lives widely, and it can be used accurately in traffic scenes is also crucial. Based on the YOLOX object detection model, a lightweight neural network for traffic object detection is proposed. First, we propose S-PANet (Simple Path Aggregation Network) structure based on the YOLOX_S baseline model. The purpose of the S-PANet structure is to reduce the parameters number of models and the amount of floating-point computation in training. Second, we propose to add CBAM (Convolutional Block Attention Module) to the network based on the S-PANet. Adding CBAM attention mechanism can improve the detection accuracy of models, but has less computation added. It is a compensation of the simplification on feature pyramid. Final, we carry out the lightweight experiments on three kinds of models, including S, Tiny, and Nano. The proposed methods are verified on KITTI and PASCAL VOC datasets. Our S model can achieve 92.9% mAP@0.5 and 21.33 FPS on KITTI dataset. The Tiny and Nano models can achieve 84.3%, 79.4% mAP@0.5 and 33.07, 41.67 FPS respectively. Compared with the baseline models, the detection accuracies mAP@0.5 have less difference and the parameters number of the models are reduced.
A Lightweight Neural Network Model of Feature Pyramid and Attention Mechanism for Traffic Object Detection
IEEE Transactions on Intelligent Vehicles ; 9 , 2 ; 3422-3435
2024-02-01
5966229 byte
Article (Journal)
Electronic Resource
English
A DENSELY CONNECTED FEATURE PYRAMID NETWORK FOR OBJECT DETECTION
TIBKAT | 2021
|Elsevier | 2024
|DOAJ | 2024
|