Abstract: An improved M2Det detection algorithm is proposed to solve the problems of low accuracy and large amount of parameters in the detection of ground arrow marking lines. The algorithm uses an improved backbone feature extraction network and a multi-level pyramid network in feature extraction, and uses non-maximum suppression to filter the generated dense bounding boxes and class scores to obtain detection results. Lightweight network named MobileNet v1 is adopted to replace the VGG network in order to reduce the number of parameters. Mish activation function is used to substitute the ReLU activation function. Meanwhile, BasicRFB module is added to the MobileNet v1 network to increase the detection accuracy. Mosaic data augmentation is also introduced to enable data augmentation. Self- labeled ground arrow lines are used as the  experimental dataset, and the experimental results show that the mAP of the improved M2Det algorithm achieves 88.72%, which is about 3.9% higher than the mAP of the original M2Det algorithm, and significantly higher than the mAP of other comparison algorithms.

Key words: arrow marking line detection, M2Det, Mish activation function, Mosaic data enhancement, average accuracy