Research On Multi-task Neural Network Algorithm For Unmanned Environment Perception

Driverless cars can bring great improvements to current traffic safety and congestion problems.Since visual perception is one of the most important technologies in driverless systems,it is important to study visual perception algorithms for driverless vehicles.Currently,most of the perception algorithms require separate models for each task such as road object detection and driveable area segmentation,which not only wastes the limited hardware resources in driverless cars,but also ignores the correlation between each task of road perception.To address the above problems and combine the characteristics of road scenes,an end-to-end model design method based on multi-task feature sharing is proposed,and a multi-task model DaSNet is designed and trained and tested with two tasks of detection of seven common road objects and pixel-level segmentation of the drivable area as examples,with the main work as follows.After studying the principles of target detection and semantic segmentation,the CSPDarknet53 backbone network is adjusted,and the road object detection model is built by combining PAN feature fusion network and YOLO detection head;the middle feature layer of the backbone network and fusion network is used as input,and the input feature layer is upsampled and fused in the second half of the network to build the second branch of the model DaSNet in this paper,that is the driveable area segmentation network.The BDD100 K dataset is extracted and preprocessed to a format matching the training DaSNet model;the advantages and disadvantages of current data augmentation methods and loss functions in the field of target detection and semantic segmentation for road scene perception tasks are analyzed,and the training data are augmented using the Mosaic data augmentation method,and the GIo U and Dice loss functions are selected to train the DaSNet multi-task model.To evaluate the performance of the DaSNet model,the YOLOv5 s,Faster R-CNN and U-Net models were trained using the BDD100 K dataset,and the performance indexes such as m AP,Dice coefficient and detection speed were compared and analyzed;further,the backbone network was replaced with CSPRes Ne Xt50 and the loss function was replaced with MSE loss and Io U loss for training,and the effects of the above three replacements on m AP and detection speed were analyzed.The results show that DaSNet multitasking model has 0.5% and 4.2% higher m AP than YOLOv5 s and Faster RCNN respectively in road object detection task,and reaches121 FPS detection speed on RTX2080 Ti GPU;CSPDarknet53 backbone network with GIo U loss has better detection speed than other backbone networks and loss functions.It has better performance in the road scene perception task;the Dice value of the segmentation branch network on the travelable area with and without priority is 4.4% and6.8% higher than that of the U-Net network,respectively,which has a more obvious improvement.Therefore,the multi-task fusion network construction method proposed in this paper effectively saves the limited computing power of driverless cars and improves the operational efficiency of the perception system in a positive way.