Real-time traffic object detection algorithm with deep stochastic configuration networks

Abstract

In computer vision and intelligent transportation, object detection algorithms are a major research hotspot for improving the perception of autonomous vehicles. Although deep learning-based object identification algorithms perform well in terms of traffic object detection, they have low accuracy in complex road settings and poor real-time performance. In order to address these issues, this paper proposes a real-time traffic object recognition technique, namely the Toward Our Dream (TOD)-You Only Look Once version 7 (YOLOv7) method, that makes use of a lightweight network model with an improved Deep Stochastic Configuration Networks (DeepSCN). Firstly, the Dilatation Mingle (DM)-Spatial Pyramid Pooling Cross-Stage Partial Convolution (SPPCSPC) module boosts object identification performance for multi-scale objects, unifies semantic information from several size feature maps, and improves the network architecture of the YOLOv7 algorithm. Secondly, an improved DeepSCN is proposed, which improves the classification performance of the detecting head. Lastly, we carried out experiments on ablation, comparison, and visual validation. The experimental findings show that our lightweight technique is better than state-of-the-art methods in terms of accuracy and real-time performance for object detection.