Abstract: Abstract: It is the key to prevent and control dairy cow diseases by real-time monitoring and detection of abnormal behavior early. In this field, it is the most common for the detection and recognition of lame behavior especially. However, the technology of the livestock abnormal behavior recognition is still facing the problem of poor real-time performance. Regarding the above problems, 2 improvement strategies were proposed in this paper. Firstly, a sparse overcomplete dictionary learning algorithm based on conjugate gradient pursuit-KSVD (CGP-KSVD) algorithm was proposed which was used to description and representation of lame behavioral semantic features. The proposed idea was to optimize search direction in the stage of sparse coding construction though introducing a conjugate gradient pursuit algorithm. Meanwhile, the proposed method could avoid the computational load caused by storing and calculating large-scale Hessen matrix effectively. By this way, it achieved the high convergence speed of the sparse overcomplete dictionary learning. Secondly, the basic features of representing the videos were extracted by the fusion of spatio-temporal interest points and spatio-temporal interest points extracted in the dense trajectory map. With the proposed method, it is possible to reduce the redundancy features and the computational load while retaining the rich details. To verify the proposed algorithm, we designed 3 experiments. Experiment 1 was used to achieve the recognition accuracy by comparing the proposed algorithm with the classical sparse dictionary learning algorithms. Experiment 2 and Experiment 3 were used to study the efficiency of the algorithm. In the experiments, 500 black and white adult cows were selected randomly and a total of 1200 video samples were achieved for training and testing by the manual marks. In Experiment 1, the average accuracy of the algorithm proposed in this paper is 5.5 percentage points higher than that of LLC (locality-constrained linear coding) and 3.9 percentage points higher than that of K-SVD. The fusion feature proposed in this paper was used as the basic features of LLC and K-SVD and the above methods were recorded as LLC + fusion feature and K-SVD + fusion feature. The experiment results show that the average accuracy of the proposed algorithm is 3.8 and 1.4 percentage points higher than that of LLC + fusion feature and K-SVD + fusion feature, respectively. In Experiment 2 and 3, the theoretical analysis results show that the algorithm proposed in this paper has the lowest computational time complexity, and the average response time of the CGP-KSVD algorithm is 0.043 s, faster than the 2 other algorithms, BP-KSVD (basis pursuit-KSVD) and OMP-KSVD (orthogonal matching pursuit-KSVD). And the 90-hour video test results show that the CGP-KSVD algorithm has the highest recognition rates of 93.3% and 92.7% respectively with the playback video and online test video, which are increased by 33.3 and 35.6 percentage points compared with the BP-KSVD algorithm, and 13.3 and 14.1 percentage points compared with the OMP-KSVD algorithm.