Aiming at the problem of low pig target detection accuracy in the complex environment in the current intelligent breeding of group-raised pigs, a lightweight target detection model for group-raised pigs based on improved YOLOX, Ghost-YOLOX-BiFPN was proposed. The Ghost convolution was used to replace the traditional convolution, which greatly reduced the number of model parameters. BiFPN was used as the model feature fusion network to effectively fuse the feature maps of pigs of different sizes, and Focal Loss function was added in the post-processing stage, increasing the learning of the model to the positive sample target, and reducing the rate of missed detection. The results showed that the improved model had a detection accuracy of 95.80% for pigs, and the number of model parameters were 2.001×107. Compared with the original YOLOX algorithm, the detection accuracy and recall were increased by 2.84 percentage points and 3.22 percentage points, respectively, and the number of model parameters were reduced by 63%. Finally, the proposed algorithm model was deployed to the Nvidia Jetson Nano mobile terminal development board. The actual operation on the development board showed that the model proposed can guarantee the recognition rate of pigs and realize the accurate recognition of pigs of different sizes and breeds. The research result can provide support for the subsequent establishment of intelligent pig breeding system.