In order to detect pigs accurately and quickly in complex environments, a multitarget pigs detection algorithm based on improved convolutional neural network (CNN) was proposed. Twolevel linear SVM was trained to generate highquality candidate regions by using binarized normed gradients (BING) of pig images. The improved CNN model was used to classify and identify candidate regions. Finally, the nonmaximum suppression (NMS) algorithm was used to eliminate redundant windows. The proposed algorithm reduced the number of training samples and parameters. Through the experiment of CNN network structure and parameter optimization, the efficiency of network training and the effect of target detection were analyzed. Experiments showed that compared with the traditional CNN model, the improved CNN model had shorter training time, faster convergence speed and stronger robustness. The classification accuracy of foreground and background of pig images was 96%, which was higher than 72.29% of the traditional CNN model. Through the analysis of false detection rate, missed detection rate and average detection time, the detection performance of this algorithm was slightly better than Faster RCNN and Yolo algorithm. The average success rate of pig tracking based on this detection algorithm was 89.17%, and the average error of center point was 6.94 pixels, which showed the effectiveness and stability of the detection algorithm in pig tracking. Using this detection algorithm, it can lay a foundation for the future research on extracting the motion parameters of pigs to judge the health status of pigs.