Object Detection And Segmentation Based On Improved Faster R-CNN In Autonomous Driving Scenarios

To improve the overall efficiency,convenience and driving safety of roads and transportation systems,autonomous driving technologies have become a focus of research and development in various countries.As one of the core technologies of autonomous driving,environmental perception uses sensors such as cameras and radar to collect data about the vehicle’s surrounding environment and assist the autonomous driving system in completing the identification and detection of road targets.Road targets are a class of targets that are significant indicators of driving vehicle behavior,such as cars on the road,pedestrians,lane lines,drivable areas,etc.ECU receives the detection results of the environment perception module and sends instructions to the vehicle control system after processing.It improves the stability of vehicle driving and the reaction speed to respond to unexpected events,helping people to drive vehicles more safely.In this paper,the target detection algorithm is improved according to the characteristics of various scales of targets in road scene.Combining traditional image processing with deep learning methods,a target detection scheme was used to detect road rectangular box targets,lane lines(curve annotation)and drivable areas(semantic segmentation annotation)to achieve a unified model for the detection of different types of road targets.The detailed research work is as follows:(1)For the morphological characteristics of the drivable area and the lane line,the solutions of area rectangle filling and curve rectangle fitting are designed respectively.For area annotation,the pixel point P in the area is located by finding the 64-dimensional connection point of the segmented annotation map and extending from P towards the area boundary to generate an extended rectangle.After masking the generated rectangle,the connected points are found in the remaining area and the rectangles are iteratively generated to fill the drivable area.For curve annotation,the position of the lane line is annotated using the outer rectangle of the curve after positioning it to ensure that the lane line is within the border annotation.(2)The Faster R-CNN target detection algorithm is improved on the problem of varying sizes and scales of border annotated targets(pedestrians,vehicles,traffic signs,etc.)in driving scenes.The area and aspect ratio distributions of the dataset annotations are analysed to set up anchor frames of suitable size,to improve the matching degree between anchor frames and targets,and thus to improve the detection accuracy of the model for tiny road targets.A feature pyramid structure is added to the feature extraction part of the model to fuse the feature maps of adjacent layers.The RPN of the model detects targets on the road at various scales based on the information from the fused feature maps at different scales.Through experimental comparison,the improved model significantly improves the detection accuracy with little increase in detection elapsed time.(3)A target detection algorithm is used to achieve drivable area segmentation.Firstly,several target boxes are used to fill the drivable area segmentation annotation,and the location of the drivable area is trained and predicted using the method based on deep learning,and its feasibility is verified by comparing the prediction effect with the semantic segmentation algorithm.And use edge extraction,contour detection and other methods to do edge smoothing process on the region border prediction results,remove the detection box in the region beyond the road boundary.Subsequently,the lane line border annotation is trained using the object detection algorithm to achieve recognition of lane line locations.Using threshold segmentation to extract the lane line picture information,compare the change in the mean grey value of the vertical area in the center of the picture before and after rotation to judge the tilt direction of the lane line,further clarify the lane line position,and finally combine the lane line information to correct the drivable area segmentation results.The experiment proves that the area edge smoothing processing and lane line correction improve the detection accuracy of the target detection algorithm for the drivable area.(4)A unified model for road target detection and drivable area segmentation.The unified model is used to train three types of annotated targets and achieve single-model detection of road border annotated targets,lane lines and drivable areas.The experimental results show that the unified model has lower hardware resource consumption and faster training and prediction speed than multiple models completing sub-tasks independently,and can better meet the real-time requirements for multiple target detection in autonomous driving scenarios.