Mental health has become a global problem, as over 300 million people worldwide suffer from depression and 200 million from anxiety disorders, which are ranked by the World Health Organization (WHO) as the first and sixth leading causes of disability, respectively. Due to the limited health resources, the traditional method of mental health diagnosis as one-to-one consultation is difficult to meet the needs of the large number of mental subhealth population. In this article, we propose a new method for mental health recognition that could identify potentially clinically significant symptoms of depression and anxiety based on daily gait. Eighty-eight participants were recruited, and their gaits were recorded by a digital camera. Then they were required to complete two rating scales, the Patient Health Questionnaire (PHQ-9) and the seven-item Generalized Anxiety Disorder Scale (GAD-7), to measure their depression and anxiety levels. Specifically, 18 key points of each individual's body trunk were captured from video, and both time-domain features and frequency-domain behavioral features were extracted for each key point. Lastly, machine-learning algorithms were utilized to build the mental health recognition models. Results showed that the proposed method is feasible and effective, with a correlation coefficient of depression (measured by PHQ-9) recognition above 0.5 and anxiety (measured by GAD-7) recognition above 0.4, achieving medium correlation. This new, low-cost, and convenient mental health recognition pattern could be applied in daily monitoring of mental health and large-scale preliminary screening of mental diseases.