Investigate the target detection algorithm and implement the development of an urban management violation detection system based on TensorFlow. By modifying the model, the map of the algorithm is increased by nearly 20%, and the algorithm detection rate is increased by 10 times. Based on TensorFlow serving, the secondary development completes the algorithm's micro-service. Tensorflow serving supports multiple GPU operations and encapsulates related java and python SDKs to facilitate the docking of other systems. Investigating and using adversarial training enhances the robustness of the model and enables the model to better adapt to different cameras. Developed a deep learning labeling aid tool, using existing models to label unlabeled data, speeding up model labeling efficiency and reducing labeling costs. Based on SORT, a demo of urban management target tracking was developed to achieve multi-target tracking. Responsible for setting up a deep learning pipeline and R & D environment. Tensorflow serving secondary development.