Research On Lane Detection Algorithm Based On Deformable Convolution And Efficient Temporal Information Processing

As one of the basic modules of the driving system,lane detection is aimed at providing the driving system with basic road information input and surrounding environment perception capability through sensor devices and lane detection algorithms.It is an important basis for the driving system to make decisions on driving direction and speed.Therefore,in situations where the vehicle speed is fast and the road environment is complex and changeable,the lane detection module needs to maintain high detection accuracy and have real-time detection speed.However,in models that target static image inputs,previous research often finds it difficult to meet these two requirements simultaneously.The "image row classification modeling method" has achieved extremely high detection speed.However,models using this modeling method have problems with decreased detection accuracy in complex scenes such as strong glare,large area obstruction,and missing lanes.In addition,although previous research on models targeting temporal video inputs can achieve high accuracy by combining temporal information,the complex processing methods used have led to extremely low detection speed.Therefore,this thesis focuses on the following research on different problems of the previous studies in the two datasets:(1)In response to the problem of decreased detection accuracy of lane detection models in visual extreme scenes,this thesis proposes a lane detection model(DefSALane)based on self-attention and deformable convolution using the image row classification modeling method.The model designs a self-attention module for high semantic features to enhance the model’s understanding of global features and focus more on the semantic features of the lane,thereby suppressing the negative effects of complex scenes such as obstruction in the image.In addition,the model combines a deformable convolution module to perform additional convolution calculations on the spatial position of the lane,thereby strengthening the use of contextual information of the lane by the model.Finally,this thesis demonstrates the effectiveness of each module of the model and proves the overall effectiveness and high computational efficiency of the model through experiments on the single-frame lane image datasets: CULane and TuSimple.(2)In response to the problem of low efficiency in utilizing temporal video information in lane models,this thesis proposes a lane detection model(RTS-Lane)based on efficient temporal information processing using the image row classification modeling method.The model uses cross-attention mechanisms to design memory generation and memory extraction modules,and combines self-attention methods to suppress noise in memory information,effectively utilizing temporal information.In addition,the model compresses the parameter amount through a mixture of Gaussian model clustering methods,reducing model calculation time.Finally,this thesis demonstrates the effectiveness of the overall model and each module through experiments on the temporal lane dataset VIL-100.(3)In response to the application scenario of road abnormal object inspection tasks,this thesis designs and develops a high-performance lane backend service process that supports asynchronous calls based on the proposed RTS-Lane model.The functionality and performance of the module were tested,and the results show that the module runs stably,meets the real-time requirements of road inspection tasks,and has better performance than other design methods.