Three-Dimensional Object Detection Based on Deep Learning in Enclosed Scenario

In recent years, due to its strong plasticity and excellent future potential, the development of automated vehicles in the mining environment is extraordinarily rapid. The application of lidar in automated vehicles has also become more and more popular, and algorithms for point cloud object detection have emerged endlessly. However, due to rough road and dusk-to-dawn operations in the mining scenario and the large size of the truck, traditional detection algorithms fail to meet the requirements of real-time updates. Due to the high precision and efficiency of the deep learning network, its application could break through the limitations of traditional algorithms. In this paper, the algorithm called PointPillars is adapted for object detection in the mining scenario. After converting the point cloud to a sparse pseudo-image and extracting features by a two-dimensional convolutional neural network (2D CNN), the time consumption of the entire algorithm becomes much less. The application of a Single-Shot Detector (SSD) detection head also improves the accuracy of the detection results. The mean Average Precision (mAP) is up to 50%, which is sufficient for application in the actual scenario. In addition, the algorithm accomplishes object detection in 69 ms, which can meet the speed requirement of lidar update and achieve the real-time application.