With the development of deep learning in the field of target detection, there are more and more types of algorithms with higher accuracy. Different from the traditional anchor_based algorithm, CenterNet proposed in recent years is an anchor_based algorithm which has the advantages of high accuracy, simple network and fast detection speed. Although CenterNet is already light and its accuracy can be maintained at a relatively high degree, there is still a space for further improvement. In order to get a better detection effect, IBN_Net normalization method and group convolution are introduced into the residual module of CenterNet to improve the accuracy of network detection and reduce parameters. In addition, Mosaic data enhancement method is also used to optimize the training mode of the algorithm to enrich the detection background and optimize the detection performance. Compared with the original network, the MAP has been improved by nearly 3% and the number of parameters is also 30% lower.