Research On Occlusion Small Target Detection Algorithm In Complex Road Scenes Based On Deep Learning

With the continuous in-depth research of intelligent driving technology for automobiles,complex road scenes object detection technology based on deep learning has been developed rapidly in recent years.However,due to the influence of many factors,such as significant changes in object scale,occlusion,complex backgrounds,changes in light intensity,and complex weather conditions,the model’s detection and recognition have a high rate of missed detection and false detection,which makes it challenging to meet the needs of detection tasks.The algorithm model at this stage has problems such as weak adaptability to complex scenes and poor robustness,and the model needs to be lighter.Given the above problems,based on the deep learning object detection algorithm,this thesis proposed an occlusion small target detection algorithm that can adapt to complex road scenes.The results show that the improved algorithm has better detection and recognition performance while balancing the model’s computational efficiency and hardware cost.The main research work of the following three parts :(1)Aiming at the problem that most of the object detection datasets of traffic scenes are foreign open source datasets and there is a lack of domestic public datasets,a new complex road scenes CQTransport dataset is constructed based on the vehicle driving recorder,including 18091 image samples.In the process of dataset construction,data sampling,strict screening,and labeling are carried out to ensure the improvement of the quantity and quality of the dataset.The data enhancement of the image solves the problem of sample imbalance,makes the CQTransport scene diversified,and improves the generalization performance of the algorithm model during the training process.(2)An improved algorithm based on an adaptive feature fusion mechanism is proposed to aim at the problem of high missed detection rate and false detection rate of the occluded object and small and medium objects in complex road scenes.The improved adjacent scale feature effective fusion module is integrated into the YOLOv5 s benchmark model,which alleviates the negative impact of the model feature fusion process.A multiscale wide receptive field adaptive fusion module is proposed to enhance the effective extraction and utilization of context information.The detection performance of the model is improved by integrating the attention mechanism,improving loss function,and increasing the prediction scale,which enhances the problem of high missed detection rate and false detection rate of small and medium objects and occluded objects in road scenes.The experimental results of multiple datasets show that the proposed improved method effectively improves the detection accuracy of small and medium objects in complex road scenes and has good robustness.The improved algorithm improves m AP by 6.7%,4.9%,and 7.9% on BDD100 K,Udacity,and CQTransport datasets.(3)The detection algorithm’s multi-scale lightweight road scene object is proposed to aim at the deployment requirements of current environmental awareness algorithms in mobile devices and embedded devices.The YOLOv5 s and YOLOX-s algorithms are benchmark models for lightweight improvement.Firstly,the Ghost Net module is integrated into the backbone network to reduce the number of parameters and calculations.An improvement strategy is further proposed to address the problem of insufficient feature extraction of the backbone network after integrating the Ghost Net module.The model’s performance is improved by adding efficient methods such as Vo VGSCSP,Sim AM,ODConv,and SPD-Conv to the network model.The experimental results on multiple data sets show that the detection speed of the improved algorithm reaches 126.6frames/s and 117.6 frames/s,respectively,which can better meet the configuration conditions of mobile devices.At the same time,it improves the detection effect of small and medium-sized targets and occlusion targets in complex road scenes and effectively solves the problems of difficult deployment and low performance of intelligent driving environment perception algorithms.