Abstract: Because of the structural complexity of the network currently available, low intensity of illumination at coal preparation plant′s raw coal handling points and underground raw coal conveyor points, the methods currently applied for detection of foreign objects on belt conveyor can hardly meet the requirements under actual working conditions in both accuracy and efficiency. To deal with this situation, an improved YOLOv8 detection algorithm that integrates the lightweight networks YOLOv8-MobileNetV1 is developed. In the YOLOv8-based model, the traditional C2F convolutional layer is replaced by lightweight network MobileNetV1 for reducing the number of parameters, and the traditional spatial pyramid pooling is replaced by large kernel pyramid pooling for further improving the model′s performance and generalization capability. Through integration with CVH attention mechanism module, the network is enabled to have a higher high-level information access capability for improving foreign objects detection accuracy and efficiency. In order to verify the effectiveness of the detection model, the database of large pieces of foreign objects and anchor bolts commonly found on belt conveyor is created, which is randomly divided at a scale of 8:1:1 into training set, verification set and testing set, for assessing the effectiveness of the proposed method in terms of detection accuracy and speed. As indicated by results of experiments, for the detection of large pieces of objects and anchor bolts with YOLOv8-MobileNetV1 algorithm, the detection accuracies are 81.30% and 89.46%, respectively, reaching an average accuracy of 85.38% and a frame rate (frames per second) of 83.5; compared with traditional detection algorithm, the use of YOLOv8-MobileNetV1 has a much higher detection accuracy and efficiency and can better meet the requirements on correctness and timeliness of detection under working conditions; and the use of the detection method can ensure safe operation of underground mine and coal preparation plant.