Under the background of urban security, the security problem of intelligent parks with dense human and vehicle has attracted increasing attention. Real-time detection of people and vehicles entering the park can provide important data reference for formulating reasonable evacuation strategies for people and vehicles in the park under emergency scenarios. In order to solve the accuracy problems caused by the complex environment, dense objects and occlusion of the current object detection during the occurrence of emergency events, as well as the rapidity problems caused by the amount of model calculation, a hybrid object detection algorithm based on improved lightweight YOLOv5 was proposed. Based on the original YOLOv5 model, the spatial pyramid pool structure SPPF in the backbone network is improved to SimSPPF to improve the real-time performance of the model. By adding an extra edge to the mesoscale layer and introducing the channel attention mechanism CA, the accuracy of the model in detecting the hybrid scene is improved. The experimental results show that compared with YOLOv5s, the algorithm not only keeps the detection speed at 142 FPS, but also improves the accuracy by 2.1 percentage points, which meets the accuracy and real-time requirements of the hybrid dynamic detection of people and vehicles in the smart park.