Cow individual recognition is the premise of automatic cow behavior analysis and disease detection，which is important for achieving precision animal husbandry. An individual identification method of dairy cows under unrestricted conditions based on the fusion of deep features and traditional features was proposed. Firstly, Mask R-CNN was used to identify cows in standing and lying positions. Secondly, two methods were used to extract the feature probability vectors of dairy cows. Convolutional neural network (CNN) was used to extract the deep features in the form of probability vectors of Softmax layer. The traditional features were manually extracted and selected by neighbourhood component analysis (NCA), and input into the support vector machine (SVM) model to output the probability vector. Finally, the two features were fused. Based on the fused features, SVM was used to classify the dairy cows. The experiment of cow individual identification was carried out on the image data set of 58 cows in standing and lying positions. The results showed that for cows in standing and lying cows, the feature fusion method improved the accuracy by about 3 percentage points and 2 percentage points compared with that using deep features alone, and the accuracy of the feature fusion method was improved by about 5 percentage points and 10 percentage points for cows in standing and lying postures, respectively, compared with traditional features alone. The accuracy of the method proposed reached 98.66% and 94.06% for standing and lying cows, respectively. The results can provide effective technical support for intelligent cow behavior analysis, disease detection, etc.