Lightweight Driving Behavior Detection Algorithm Combined with YOLOv5 and Pruning Distillation

Irregular driving behavior exhibited by drivers could result in road traffic accidents, posing significant risks and potential dangers. The implementation of real-time detection systems for identifying irregular driving behavior emerges as a promising approach to address the problem effectively. In...

Full description

Saved in:
Bibliographic Details
Published in:Chinese Automation Congress (Online) pp. 1130 - 1135
Main Authors: Zou, Peng, Yang, Kaijun, Zhang, Jianqiang
Format: Conference Proceeding
Language:English
Published: IEEE 17.11.2023
Subjects:
algorithm Deployment
Automation
channel pruning
Detection algorithms
Image edge detection
improper driving behavior
knowledge distillation
Performance evaluation
real-time detection
Real-time systems
Road traffic
YOLO
ISSN:2688-0938
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Irregular driving behavior exhibited by drivers could result in road traffic accidents, posing significant risks and potential dangers. The implementation of real-time detection systems for identifying irregular driving behavior emerges as a promising approach to address the problem effectively. In this paper, We design a light-weight irregular driving behavior detection algorithm combined with pruning distillation algorithm and integrate the proposed algorithm into the YOLOv5 model. Knowledge distillation is applied to the pruned model, facilitating fine-tuning and enhancing the accuracy of the model's detection capabilities for irregular driving behavior. To validate the feasibility of our algorithm in detecting irregular driving behavior, we conducted object detection experiments using three models: the original YOLOv5n algorithm, the pruned YOLOv5n-finetune-pruned, and the fused channel pruning and knowledge distillation YOLOv5n-distillation. The results demonstrate that the improved YOLOv5n-distillation algorithm achieved the remarkable P (98.6%) and R (98.9%) values, while reducing the model size by 20%. This reduction enables efficient deployment on small edge devices, without compromising the accuracy of detecting irregular driving behavior.
ISSN:2688-0938
DOI:10.1109/CAC59555.2023.10452015