Research On Muti-frame Instance Lane Detection Algorithm Based On Deep Learing

Road safety,as a significant topic in the field of vehicles,has received widespread attention in academia and industry.In practical applications,video-based lane detection technology encounters challenges due to interframe evolution in dynamic scenarios.Currently,in the field of deep learning,the inter-frame evolution issue in image-level lane detection methods primarily manifests as a lack of frame-to-frame correlation,resulting in weak model adaptability.In video-level lane detection methods,the inter-frame evolution issue leads to weak frame-to-frame correlation and consequently lower detection accuracy in complex scenarios.Existing lane detection methods based on video object segmentation suffer from drawbacks such as poor real-time performance and weak inter-frame correlation.Therefore,this paper aims to address the limitations of videobased lane detection in terms of accuracy and stability by focusing on researching road perception methods in dynamic scenarios.The specific work includes the following:(1)To balance the accuracy and inference speed of existing models,the paper proposes a fast video instance-based lane detection network.The model first utilizes a lightweight backbone network as the feature extraction part for two encoders.Secondly,a fusion and attention module is introduced in the neck of the memory frame encoder to provide more specific spatial positions and instance queue information for each instance of lane markings.Lastly,a global context module is proposed in the neck of the query frame encoder,which can be used for model decoding to enhance the position information of the current frame across multiple scale features.Experimental results on the VIL-100 dataset demonstrate that compared to the baseline network,this network not only achieves higher detection accuracy but also improves the inference speed by 2.27 times.(2)To address the issue of inter-frame evolution in existing models,the paper proposes a video instance-based lane detection network using memory templates.This model is based on a lightweight architecture for video instance-based lane detection.By establishing a strong correlation between previous frames and the current frame using memory templates,the model captures global and local dynamic lane features to handle lane occlusion and inter-frame evolution issues caused by dynamic environments.Furthermore,to address the inherent error problem of the spatiotemporal memory network,the paper utilizes the Grad-CAM technique to locate and visualize the inherent errors in the memory.In order to mitigate the impact of such errors on network accuracy,the paper improves the generation of transfer matrices,which are used to independently calculate the temporal consistency loss for each instance lane.The proposed model was validated on both image-level and videolevel datasets.The results from dataset experiments and real-world vehicle experiments indicate that the model outperforms the state-of-the-art videolevel instance-based lane detection algorithms when dealing with dynamic traffic scenes.The proposed model maintains fast detection while improving the stability of detection results over time,making it more robust in practical dynamic traffic scenarios.