Lightweight Landmark Perception Algorithm And Its Application In Cognitive MAP

As an important part of intelligent driving environment perception and cognitive map semantic layer,landmarks can not only provide necessary traffic information,but also assist in positioning and planning,controlling and decision-making,which participate in the analysis to provide the corresponding positioning results and decisionmaking schemes to control the intelligent vehicle to operate correctly.Therefore,the research on landmark recognition has important theoretical and practical significance.In order to recognize landmark information efficiently,based on the visual information collected from the driving environment by the camera,this paper studies the recognition algorithm for significant landmarks including traffic signs,buildings,street lights.conducts relevant experiments and analysis based on deep learning algorithm.And based on deep learning algorithm,it carries out relevant experiments and analysis.The main contents are as follows:1.Aiming at the distribution characteristics of landmarks and the change of target size in real driving scene,this paper proposes a landmark perception algorithm based on hyper feature pyramid by using convolutional neural network.Through the organic combination of deep features and shallow features,a more computationally friendly hyper feature pyramid network is proposed which is based on the feature pyramid network.Multi-scale feature information is integrated to extract more representative,multi-level and multi-scale features.In addition,this paper also designs a new improved feature extraction network and integrates hyper feature pyramid network into it,using the Landmark Integrated Database for training and testing.The result shows that the network model designed in this paper achieves mAP@0.5 of 82.1736%,which is 10.578% higher than that of YOLOv3,the most widely used object recognition network in both academia and industry,but our model is smaller.2.In view of the high requirement of complex deep model for the hardware of actual intelligent driving condition,based on attention mechanism,a lightweight algorithm for the above landmark perception model is proposed.With the help of attentional mechanism,the algorithm explicitly models the dependency relationship between feature channels and adaptively obtains the importance of each feature channel.Through the importance of the channel layer of the feature map,the corresponding filters of the upper layer and the corresponding convolutional kernel of the filters of the next layer are pruned.Results show that after pruning the landmark recognition network,the model is reduced by about 75%.The precision is similar to the original model,and the mAP@0.5 can still reach 81.4%,effectively eliminating redundant parameters.In order to verify the generalization ability of the proposed perception algorithm,this paper conducts a real vehicle test oriented to cognitive map based on the intelligent vehicle driving platform.Besides,this paper introduces the hardware and software of the platform,and completes the performance test under the driving scene of the campus.Experimental results verify the robustness of the proposed lightweight landmark perception algorithm.