Research On Road Obstacle Detection Algorithm Based On Feature Fusion And Attention Mechanism

With the continuous development of information technology,object detection has become a basic task in the field of computer vision,and has been widely used in medical diagnosis,autonomous driving,and smart services.Among various applications based on target detection,autonomous driving has become the main research direction of related scholars in recent years because of its convenience and scalability,and road obstacle detection is an important part of it.Road obstacle detection based on deep learning can achieve high detection accuracy and fast detection speed,so it has extremely important research significance and practical application value.However,the current mainstream road obstacle detection algorithms have the problem of low detection accuracy for small targets or partially occluded targets in practical applications,which cannot meet people’s high safety requirements for autonomous driving.In view of the existing problems,this paper proposes a method that can accurately detect the target and can be applied to the actual scene according to the existing technology.The main work of this paper is as follows:Based on the target detection algorithm Faster R-CNN,this paper firstly adds a feature pyramid network FPN that fuses deep and shallow features from top to bottom in the Faster R-CNN network,that is,it uses deep features with high semantic information at the same time.and high-resolution shallow features,so that it can process the position and scale-transformed objects.Then,in view of the feature contradiction between the context information and the target image after the Faster RCNN algorithm is combined with FPN,a feature pyramid transformation FPT based on the attention mechanism is introduced,so that the feature pyramid can increase the spatial semantic information of the target feature.Finally,the features of the two parts of FPN and FPT are fused,so that the model can extract the feature map through the interaction of space and scale,that is,the context information and the feature information of the target are fused,so that the improved algorithm can improve the detection accuracy of small target or partially occluded objects.Different feature information after fusion will have different effects on the detection result,and some feature information is invalid for the detection result,and even leads to the degradation of the model performance.In response to this problem,this paper introduces a channel attention mechanism to strengthen the feature fusion part,that is,selectively enhances important features by adaptively adjusting the weights of different channels,so that the algorithm model can improve the discrimination of important features,thereby improving the accuracy of the detection performance of small targets or partially occluded targets,while improving the convergence of the model.In this paper,experimental comparison and visual analysis of the proposed methods are carried out on PASCAL VOC,KITTI and self-made indoor datasets,respectively,and the detection effect in actual road scenes is verified by designing system modules.The experimental results show that the improved method The experimental results show that the improved method based on the ResNet-101 network improves the detection accuracy of small targets by 5.29 percentage points in the KITII dataset,which also proves that the improved obstacle detection method is suitable for practical application scenarios.It can provide information about the surrounding road conditions for vehicles to avoid obstacles as soon as possible,which greatly improves the safety.