Abstract: Traditional object detection model uses artificial object features, resulting in poor detection accuracy. Object detection model based on deep learning has high detection accuracy. However, for application of coal mine rescue robot with high real -time and accuracy requirements, the obtained image information is less and object features are not obvious, resulting in poor object detection effect. In order to improve accuracy and speed of object detection, on the basis of YOLO V3 model, an object detection model of coal mine rescue robot based on multi -scale feature fusion is proposed. The model mainly includes feature extraction module and feature fusion module. The feature extraction module uses hole bottleneck and multi -scale convolution to obtain more abundant image feature information, so as to enhance expression ability of object feature and improve object classification accuracy and detection speed. The feature fusion module introduces spatial attention mechanism into feature pyramid to effectively fuse high -level feature map with rich semantic information and low -level feature map with rich location information, which makes up for lack of position information expression ability of high -level feature map, and improves object positioning accuracy. The model is deployed on embedded NVIDIA Jetson TX2 platform in coal mine rescue robot for post -disaster environment object detection experiment. The detection accuracy is 88.73% and detection speed is 28 frames per second, which meet real -time and precision requirements of object detection of coal mine rescue robot.