Research On Target Detection Algorithm Based On Monocular Vision

In recent years,object detection technology centered around monocular vision has received widespread attention from both industry and academia,and has been applied in areas such as intelligent monitoring,human-computer interaction,and autonomous driving.With the development of deep learning,data-driven convolutional neural networks have been used to identify and locate regions of interest in images.A review of current research on object detection reveals that while significant progress has been made in some specialized tasks such as face recognition and industrial quality inspection,existing approaches face performance bottlenecks in handling objects with varying scales,diverse types and quantities,limited visible pixels,and a single imaging perspective,due to insufficient feature extraction,limited receptive fields,and imbalanced positive and negative samples.Therefore,researching high-precision,high-robustness,and high-efficiency object detection is a key task in monocular vision,and also represents the transformation from academic validation research to practical applications in industry.This paper analyzes several key issues in current object detection,summarizes the development path and research status of object detection.Based on this,the research focuses on the selection of backbone networks,the construction of receptive field enhancement modules,the design of self-adaptive feature recombination and fusion modules,the embedding of self-calibrating convolutions,and the division of training samples.The main research contents are as follows:(1)Aiming at the problems of insufficient detection accuracy and poor detection performance for small objects in the baseline algorithm FCOS(Fully Convolutional One-Stage Object Detection),a new FCOS object detection algorithm based on multi-branch atrous convolution and adaptive feature fusion is proposed.Optimizing the backbone using Res Ne St and enhancing the diversity of feature representation through multi-feature map operations.A receptive field enhancement module based on multi-branch atrous convolution is constructed to expand the receptive field and obtain global contextual information.An adaptive recombination feature fusion module is designed based on content-aware recombination and attention mechanisms,which efficiently integrates the semantic information of high-level feature maps and the detail information of lowlevel feature maps.Experimental results on multiple public datasets show significant improvements in accuracy compared to the baseline algorithm.(2)Aiming at the problems of imperfect bounding box regression and suboptimal positive and negative sample division in the baseline algorithm,a new RA-FCOS object detection algorithm based on adaptive training sample division strategy is proposed.The Res Ne St network is optimized using self-calibrating convolution to further enhance the ability to extract representative features of targets.An adaptive training sample division strategy is designed to increase the number of positive samples during the training process and solve the problem of suboptimal positive and negative sample division.A scale-invariant Intersection over Union(Io U)strategy is introduced to optimize the bounding box regression function of the baseline algorithm and improve the accuracy of bounding box regression.Experimental results show that the proposed method can effectively improve the accuracy of object detection and outperforms other object detection algorithms on public datasets.(3)Aiming at the target detection task of vehicle,pedestrian and non-motor vehicle in real road scene,a perception system based on monocular vision is constructed,and the validity and rationality of the proposed algorithm is analyzed.An autonomous driving data acquisition platform is built,using a monocular camera as the sensor for data collection.Based on the collected images,data is augmented,enhanced,and annotated to construct an autonomous driving image dataset that covers real-world road scenarios.The proposed algorithm is trained and tested on the selfmade dataset,and the detection accuracy in actual scenarios is evaluated and compared.A visualization platform is designed to verify and analyze the real-time detection performance of the algorithm in complex traffic scenarios.