3D Position Measurement Algorithm for Military Vehicles Based on Deep Learning

In order to locate the 3D position of military vehicle in the coordinate of smart ammunition, a monocular vision measurement method using deep learning is proposed. The target is regarded as a point so that the estimation of target center is the core problem. Target detection and position measurement can be performed simultaneously. The proposed model contains feature extraction network and measurement network. To carry out long-distance measurement, the proposed method first requires feature maps with a high resolution, so the Deep Layer Aggregation (DLA) network is used as a feature extraction network. Then the idea of regression applied to the measurement network. The network heads which include two convolutional layers and a ReLU between them, regress the target center probability heatmap, 2D bounding box size, center bias, and depth. Experimental results show that when the measurement distance is less than 400 meters, the relative positioning error is less than 3%. The proposed algorithm has the advantage of low complexity and does not require prior information such as geometric features, target sizes and can achieve end-to-end single-shot target depth prediction.