Multi-task Deep Learning Method For Lane Detection And Object Detection

The deep integration of autonomous driving technology and artificial intelligence technology has a profound impact on the field of intelligent transportation.Autonomous driving technology is mainly composed of environment perception,decision planning and intelligent control,among which environmental perception is the premise and basis of decision planning and intelligent control,and it is also the most active field in artificial intelligence research.Lane detection and object detection are two of the most important tasks in the environment perception,and they are also the parts that consume more computing resources.For urban automatic driving,lane detection and object detection are mostly two independent modules,each module occupies independent computing resources,and the two modules have the same input,and the subsequent decision planning part needs to fuse the output results of the two modules before it can be carried out.The way of using two independent modules to realize lane detection and object detection tasks is not only cumbersome,occupies a lot of computing resources but also it is difficult to meet the real-time processing requirements.For this situation,the thesis summarizes the existing lane detection and object detection methods based on deep learning and conducts an in-depth analysis of their network architectures.On this basis,we try to integrate the lane detection network architecture and object detection network architecture based on deep learning,realize the simultaneous completion of lane detection and object detection with a network model,then transplant the fusion model to the embedded device to achieve real-time processing on the embedded device.The specific research in this thesis mainly contains four aspects as follows:(1)The design of the fusion network of lane detection based on semantic segmentation and object detection method is completed by sorting and experimenting related literatures of lane detection and object detection algorithm.Thesis proposes a multi-task fusion network that can simultaneously perform lane detection and object detection.The network consists of a shared encoder and two decoders,where the encoder is used to extract image features and the decoder is used to detect specific tasks.A single model is implemented to handle both lane detection and object detection tasks.(2)Completed the transplantation and deployment of multi-task network on the embedded computing platform Apex Xavier.The connections between unimportant neurons in the network are removed by optimization,and the pruned network is further compressed using quantization and weight sharing.It is deployed on the embedded computing platform based on the Tensor RT library.(3)The sorting of open source datasets and actual collected datasets are completed.In order to verify the generalization and stability of the multi-task network in different traffic scenarios,arranged the sample data of typical scenarios including highways,urban roads,congestion scenarios,rainy days,foggy days,and tunnels.(4)The performance evaluation of multi-task networks is completed on open source datasets and our datasets.The experimental results on the BDD100 K dataset show that the lane segmentation accuracy of the designed multi-task network is about68.7%,and the object detection accuracy m AP@0.5 is about 72.3%.The detection speed is about 62 FPS on the desktop computer,and the detection speed without acceleration is about 17 FPS after transplanting to the embedded computing platform.After acceleration using the Tensor RT library,it can reach 58 FPS,realizing the effect of real-time detection.