The drowning incidents are frequent in natural waters. The detection of individuals is difficult from the surveillance point of view due to the small size of the target and occlusion. In order to address this situation, a detection method for overboard individuals in natural waters based on YOLOv7-FAEA is proposed. Firstly, the YOLOv7 network is used as the baseline and FReLU is used as the activation function, which is more suitable for visual tasks. Then, ACmix attention mechanism is added to combine the self-attention mechanism with convolution, which reduces the computational overhead and increases the detection accuracy of small targets. EIoU loss function is adopted to improve the feature extraction ability of the network. Finally, the region detection module is incorporated to improve the detection efficiency. The results show that as compared with the YOLOv7 algorithm, there is a 5.1% increase in accuracy, 1.2% increase in recall rate, 4.9% increase in the mean average accuracy, and 4 increase in FPS for the proposed algorithm. YOLOv7-FAEA has an accuracy of 84.9%, a recall of 92.7%, an average accuracy of 90.2%, and an FPS of 87 on the homebrew dataset. YOLOv7-FAEA can be deployed in the monitoring server to manage the security monitoring of natural waters through cameras.