Research On Lane And Vechicle Detecion For Small Four-rotor Unmanned Aerial Vechicle

With the increasing application of UAV in urban traffic management,lane and vehicle detection as the basic and important part of UAV traffic environment perception play an important role in urban transport fields.Especially in monitoring illegal behavior of vehicle,traffic flow statistics and road automatic inspection,which is practical significance to promote the development of urban transportation.Through a lot of research and analysis at home and abroad,most of the existing methods of lane and vehicle detection are applied in the field of intelligent vehicle environment perception,and research on UAV field is less.This thesis study the problem of the lane and vehicle detection of UAV,in the aspect of lane detection,for the traditional lane detection algorithm,it is difficult to extract the features of lane and the detection accuracy is low under the condition of illumination and occlusion,in addition,the algorithm of lane detection based on deep learning has some problems,such as time consuming,therefore,a fast lane instance segmentation network based on separable convolution is proposed,it can divided lane into different instances to realize the multiple lane detection.In the aspect of vehicle detection,in order to solve the problems of low accuracy and poor robustness of existing UAVs for the small scale target vehicles,a vehicle detection algorithm based on multi-scale optimization is proposed,the sub-sampling semantic features of different scales are fused with the up-sampling features,and the various scales of vehicle are predicted to improve the detection accuracy of the vehicle.The main contents of this paper are as follows:1.The state-of-the-art algorithms for lane and vehicle detection of UAV are given,with a summary of their pros and cons.The problem of lane and vehicle detection is summarized,and the motivation of this thesis is given.2.A fast instance segmentation algorithm for lane detection is proposed.Firstly,the deep separable convolution module is used to construct the semantic segmentation network of the lane,moreover,in order to further improve the segmentation accuracy,the embedding branch network is constructed to predict the embedded spatial features of the lane pixels.Finally,the lane instance segmentation is realized by clustering algorithm.3.A vehicle detection algorithm based on multi-scale optimization is proposed.Firstly,the idea of target detection based on one-stage regression is adopted,and the anchor mechanism is introduced.In the process of feature decoding,the multi-scale optimization strategy that is respectively upsampling the feature maps of 8 times,16 times and 32 times of encoding layer is adopted,the each result are fused with the corresponding encoding layer,the network can output three different scale information,it can improve the accuracy of vehicle detection.4.The effectiveness of the proposed detection method is verified by comparing the several existing detection methods on own lane dataset and opened vehicle dataset of UAV.Finally,the experiment is performed to assess the performance of proposed method,compared other methods,the proposed method has better performance in metrics of accuracy and robustness.This thesis provided the practical significance for the following research on the field of traffic environment perception of UAV.