| Computer vision-based road vehicle object detection technology is the core technology of intelligent transportation system and intelligent vehicle system.It plays a key role in perceiving the road environment.With the fast development of deep learning technology,significant technological advancements have been achieved in vehicle object detection technology.However,high resource for computing and significant video memory consumption are required in kinds of deep learning-based vehicle object detection algorithms.It raises the cost of the algorithm model to be implemented in practice.Lightweight neural networks with fewer parameters and less computation are more suitable for the situations with constrained computing power and storage space,such as mobile terminals,embedded devices,and other edge computing devices.To address the issues that the current vehicle object detection algorithm model has many parameters,large amounts of calculation,and is difficult to deploy on mobile terminal devices,a lightweight network-based road vehicle object detection algorithm is studied and based on the theories of convolutional neural network and object detection.This thesis includes the following contents:(1)A vehicle object detection algorithm for road scenes is designed.Missing detection and false detection are common due to the numerous vehicle targets and severe occlusion in the road scenes.To address this problem and based on the YOLOv5 algorithm,a convolution module that integrates the attention mechanism is firstly designed to improve the network’s ability to extract features,then an attention mechanism is embedded in the detection head,finally the non-maximum suppression algorithm and the loss function are optimized and improved.The experimental results demonstrate that the enhanced algorithm model still improves the average precision by 2.6 percent without increasing the model’s complexity,only increasing the number of parameters by 0.1 percent and the amount of computation by 0.6 percent,which improves the algorithm model’s performance of vehicle object detection in road scenes.(2)A road vehicle object detection algorithm based on lightweight network is designed.The vehicle object detection algorithm model for road scenes is transformed to lightweight network.Firstly,a lightweight and efficient backbone network is designed by importing efficient channel attention mechanism.Then the lightweight transformation is carried out on the neck layer.Finally,the feature fusion mode of the neck layer is enhanced and optimized.Compared with the original YOLOv5 model,the experimental results display that the proposed lightweight road vehicle object detection algorithm trades off 1 percent in average accuracy,but the number of parameters is reduced by 73.2percent,the calculation is reduced by 58.2 percent,and the speed of detection is increased by 7 frames per second.Only a small amount of precision is sacrificed,but the parameters and the calculation of the network are greatly reduced.A lightweight network-based road vehicle object detection algorithm is proposed in this thesis.It minimizes the model’s computational and spatial complexity,achieves accuracy and real-time performance while being lightweight,and can still maintain robustness even in complex road scenes.The study presented in this thesis is significant both theoretically and practically,contributing to the advancement of technical levels such as the intelligent vehicle and intelligent transportation system. |
