Abstract: Target detection technology is widely used in transportation, medical treatment, security, aerospace and other fields. At present, target detection technology faces challenges such as weak target, complex background and blocked target. Aiming at the problems of low detection and recognition rate and high false recognition rate of small rock and stone targets in the images collected from unstructured simulated terrain of star catalog, the YOLOv5 target detection algorithm with the best effect and lightweight model is studied. On the basis of it, a rock target detection algorithm meeting the real time requirements is improved via improving the optimizer and optimizing the repeated detection effect of detection frame. In particular, the feature description ability and classification accuracy of YOLOv5 network model are improved by introducing empty labels as negative sample, combining SGD optimization model and nonmaximum suppression parameter adjustment method. Using complex terrain images collected in ground test sites as data sets, and using mAP, FPS and PR curves as performance indexes, the proposed target detection network is verified by experiments. The experimental results show that the improved network proposed in this paper has higher accuracy and lower false reputation rate, while maintaining the real time performance of the original algorithm.

Key words: planetary terrain, stone detection, deep learning, convolution neural network