This paper proposes an approach for automatic detection of abnormal human gait. We use an improved skeleton data covariance based gait assessment approach. Low-limbs flexion angles are derived using skeletons computed from data acquired by the Kinect sensor. Then for each gait sequence, we calculate a covariance matrix from the obtained angles data. The matrices are used as features for two classification schemes: a normal gait model-based and a k-NN-based. The resulting descriptor is compact, does not require prior temporal segmentation and shows competitive results on available pathological gait datasets.