Towards automatic detection of road features by deep learning

In Japan, a significant amount of point cloud data – a collection of data points in space – has been measured and accumulated for public works using mobile mapping systems and surface laser scanners. soil. However, this huge amount of data is of limited use in its unprocessed and unstructured state. Fortunately, it can be structured by automatically extracting an object using “as-built diagrams” that represent the complete shape of a constructed object.

Earlier this year, researchers from Japan, led by Professor Ryuichi Imai of Hosei University, Japan proposed another method to extract Road features using high-precision 3D (HD) map data. However, the applicability of their approach is limited to the developed parts of the road map. Although the problem can be solved by deep learning-based recognition, they require a large amount of hand-prepared, high-quality training data.

Recently, Professor Imai and his collaborators, Kenji Nakamura of Osaka University of Economics, Yoshinori Tsukada of Setsunan University, Noriko Aso of Dynamic Map Platform, and Jin Yamamoto of Hosei University have developed an algorithm to automatically automate training data generation and build a line feature recognition model from point cloud data automatically extracted from HD maps.

“Currently, people need to visually inspect point cloud data to identify road features because computers cannot recognize them. But with our proposed method, feature extraction is possible. is done automatically, including features at undeveloped road map segments,” explains Professor Imai. . Their work was presented at the 12th Joint International Conference on Soft Computing and Intelligent Systems and the 23rd International Symposium on Advanced Intelligent Systems on November 29, 2022. .

In their study, the researchers first separated the land surface from the point cloud data using CloudCompare software. Next, they generate the area data from the HD map and extract the component points of the feature. Although these points are designated as traffic signs or traffic lightsOther labels were provided for the remaining data.

Then, the region data corresponding to the component points is expanded to generate the training data. Using this, the researchers created additional point cloud projections. Finally, they used the training data to build a recognition model using the YOLOv3 object detection algorithm. The model can detect line features based on group points that are different from those defined for ground use CloudCompare.

After establishing the computational framework, the researchers performed demonstration experiments in Shizuoka Prefecture on a road with 65 traffic signs, 46 traffic lights, and noise characteristics within a distance of 1.5 hours. kilometer. They used 258 traffic signs and 168 traffic lights to train their recognition model and used 36 and 24 images respectively to calculate the accuracy of the identification algorithm.

The researchers found that accuracy, recall, and F-measures were 0.84, 0.75, and 0.79 for road signs, and 1.00, 0.75, and 0, respectively. 86 for traffic lights respectively, showing no misidentification. The accuracy of the proposed model is proven to be higher than that of existing models.

Professor Imai concluded by highlighting the future implications of the work. “The product model built from point cloud data will enable the realization of a digital twin for urban spaces with regularly updated road maps. It will be indispensable for management. and reduce traffic restrictions and road closures during road checks.This technology is expected to reduce time costs for people using roads, cities and other infrastructure in everyday life. their day.”

Provided by Hosei . University