Research And Implementation Of Object Detection Algorithm For Road Scene

Object detection is a key technology for real-time scene perception in autonomous driving,which plays a guiding role in vehicle central control systems.It has become a research hotspot for major internet companies and traditional car companies committed to technical innovation,and is a key core technology that combines convenience and safety.Traditional object detection cannot meet the high requirements of real-time detection for autonomous driving.In addition,due to the lack of computing power of mobile devices such as vehicles,complex urban road scenes and various weather change factors,its application has certain limitations.To address the above issues,this paper aims to improve the comprehensive performance of the model from various perspectives such as lightweight,contextual information and attention mechanism.Based on existing models,it further optimizes the object detection model for road scenes,ensuring both efficiency and detection accuracy.Finally,we designed the system based on the above algorithm model.The main work of this paper is as follows:(1)Aiming at problems such as insufficient storage space and computing resources for mobile devices such as vehicles,and considering that there are a large number of small objects in road scenes and the object scale can also change with the movement of the perspective,which can easily lead to missed detection and false detection.This paper conducts a comparative analysis of the existing lightweight backbone network Mobile Net series and Efficient Net series,the YOLO-B2(YOLO Only Look Once-B2,YOLO-B2)model based on Efficient Net-B2 was designed to reduce the amount of network params and computation,while also ensuring the accuracy.Based on the YOLO-B2 model,we propose a pyramid integration module based on contextual information enhancement to solve the problem of loss of high-level feature information before feature fusion.The module increases the model perceptual field by introducing maxi-pooling and dilated convolution to obtain rich contextual information to solve the problem of insensitive small object detection.The experimental comparison with existing excellent methods has been carried out to improve the accuracy without adding a few params and the experimental results show that our methods have better results.(2)To address the problems of information redundancy and interference of complex natural environment background factors caused by the top-down dissemination of information before feature fusion.We design an attention enhancement network based on a dual-attention mechanism to alleviate the information redundancy after feature fusion in order to improve the multi-scale feature representation of the model in a heterogeneous background environment.The proposed network integrates a channel attention mechanism as the initial step,which assigns weights to adjust the importance of each channel.It subsequently employs cross-alignment to unify the resolution of feature maps.Finally,a spatial attention mechanism is introduced to capture object locations more accurately and improve object localization in the spatial dimension.Experimental results clearly demonstrate that the utilization of the dual attention mechanism yields significant improvements in object detection and localization capabilities,ultimately leading to enhanced accuracy.(3)Finally,this paper describes the design process of the object detection system for road scenes,including the functions of each module of the system and the visualization interface.Additionally,to visually demonstrate the effectiveness of the aforementioned object detection model,this paper designs and implements an object detection system for road scenes,incorporating picture detection,video detection,and real-time camera detection functionalities.