Все выпуски

Sensor fusion is one of the important solutions for the perception problem in self-driving cars, where the main aim is to enhance the perception of the system without losing real-time performance. Therefore, it is a trade-off problem and its often observed that most models that have a high environment perception cannot perform in a real-time manner. Our article is concerned with camera and Lidar data fusion for better environment perception in self-driving cars, considering 3 main classes which are cars, cyclists and pedestrians. We fuse output from the 3D detector model that takes its input from Lidar as well as the output from the 2D detector that take its input from the camera, to give better perception output than any of them separately, ensuring that it is able to work in real-time. We addressed our problem using a 3D detector model (Complex-Yolov3) and a 2D detector model (Yolo-v3), wherein we applied the image-based fusion method that could make a fusion between Lidar and camera information with a fast and efficient late fusion technique that is discussed in detail in this article. We used the mean average precision (mAP) metric in order to evaluate our object detection model and to compare the proposed approach with them as well. At the end, we showed the results on the KITTI dataset as well as our real hardware setup, which consists of Lidar velodyne 16 and Leopard USB cameras. We used Python to develop our algorithm and then validated it on the KITTI dataset. We used ros2 along with C++ to verify the algorithm on our dataset obtained from our hardware configurations which proved that our proposed approach could give good results and work efficiently in practical situations in a real-time manner.

Sensor fusion is one of the important solutions for the perception problem in self-driving cars, where the main aim is to enhance the perception of the system without losing real-time performance. Therefore, it is a trade-off problem and its often observed that most models that have a high environment perception cannot perform in a real-time manner. Our article is concerned with camera and Lidar data fusion for better environment perception in self-driving cars, considering 3 main classes which are cars, cyclists and pedestrians. We fuse output from the 3D detector model that takes its input from Lidar as well as the output from the 2D detector that take its input from the camera, to give better perception output than any of them separately, ensuring that it is able to work in real-time. We addressed our problem using a 3D detector model (Complex-Yolov3) and a 2D detector model (Yolo-v3), wherein we applied the image-based fusion method that could make a fusion between Lidar and camera information with a fast and efficient late fusion technique that is discussed in detail in this article. We used the mean average precision (mAP) metric in order to evaluate our object detection model and to compare the proposed approach with them as well. At the end, we showed the results on the KITTI dataset as well as our real hardware setup, which consists of Lidar velodyne 16 and Leopard USB cameras. We used Python to develop our algorithm and then validated it on the KITTI dataset. We used ros2 along with C++ to verify the algorithm on our dataset obtained from our hardware configurations which proved that our proposed approach could give good results and work efficiently in practical situations in a real-time manner.

Точное определение деревьев имеет решающее значение для экологического мониторинга, оценки биоразнообразия и управления лесными ресурсами. Традиционные методы ручного обследования трудоемки и неэффективны на больших территориях. Достижения в области дистанционного зондирования, включая лидар и гиперспектральную съемку, способствуют автоматизированному и точному обнаружению в различных областях.

Тем не менее, эти технологии обычно требуют больших объемов размеченных данных и ручной инженерии признаков, что ограничивает их масштабируемость. Данное исследование предлагает новый метод самообучения (Self-Supervised Learning, SSL) с использованием архитектуры SimCLR для улучшения классификации видов деревьев на основе неразмеченных данных. Модель SSL автоматически обнаруживает сильные признаки, объединяя спектральные данные гиперспектральной съемки со структурными данными лидара, исключая необходимость ручного вмешательства.

Мы оцениваем производительность модели SSL по сравнению с традиционными классификаторами, такими как Random Forest (RF), Support Vector Machines (SVM), а также методами обучения с учителем, используя набор данных конкурса ECODSE, который включает как размеченные, так и неразмеченные образцы видов деревьев на биологической станции Ordway-Swisher во Флориде. Метод SSL показал значительно более высокую эффективность по сравнению с традиционными методами, продемонстрировав точность 97,5% по сравнению с 95,56% для Semi-SSL и 95,03% для CNN при обучении с учителем.

Эксперименты по выборке показали, что техника SSL остается эффективной при меньшем количестве размеченных данных, и модель достигает хорошей точности даже при наличии всего 20% размеченных образцов. Этот вывод демонстрирует практическое применение SSL в условиях недостаточного объема размеченных данных, таких как мониторинг лесов в больших масштабах.

Accurate tree identification is essential for ecological monitoring, biodiversity assessment, and forest management. Traditional manual survey methods are labor-intensive and ineffective over large areas. Advances in remote sensing technologies including lidar and hyperspectral imaging improve automated, exact detection in many fields.

Nevertheless, these technologies typically require extensive labeled data and manual feature engineering, which restrict scalability. This research proposes a new method of Self-Supervised Learning (SSL) with the SimCLR framework to enhance the classification of tree species using unlabelled data. SSL model automatically discovers strong features by merging the spectral data from hyperspectral data with the structural data from LiDAR, eliminating the need for manual intervention.

We evaluate the performance of the SSL model against traditional classifiers, including Random Forest (RF), Support Vector Machines (SVM), and Supervised Learning methods, using a dataset from the ECODSE competition, which comprises both labeled and unlabeled samples of tree species in Florida’s Ordway-Swisher Biological Station. The SSL method has been demonstrated to be significantly more effective than traditional methods, with a validation accuracy of 97.5% compared to 95.56% for Semi-SSL and 95.03% for CNN in Supervised Learning.

Subsampling experiments showed that the SSL technique is still effective with less labeled data, with the model achieving good accuracy even with only 20% labeled data points. This conclusion demonstrates SSL’s practical applications in circumstances with insufficient labeled data, such as large-scale forest monitoring.