We propose a deep regression-based neural network to analyze gait of individuals in low resolution video for real-time detection of abnormal activity/threats in surveillance applications. In todays commercial setting, extracting such gait patterns require motion capture devices such as Kinect or VICON, and is restricted to indoor and controlled scenarios such as in gaming applications or clinical studies. The network is trained by incorporating an inverse kinematic Groebner-based model to estimate the body joint angles from the pose. These angle trajectories of the upper and lower extremities of the body serve as gait signatures for identifying threat patterns. The first few layers model the relationship between motion and image features of the individual using a deep belief network. The next few layers model the relationship between the latent features generated from deep belief nets and the inverse kinematic model using a regression-based deep network. This network characterizes the relationship between the low-level image/motion features and the kinematics associated with the movement of an individual. The estimated joint angle trajectories are then classified as threats (person wearing a loaded vest) and non-threats (person without any load on body) using a K-Nearest Neighbor classifier. Experimental results on the INSPIRE dataset released by Air force institute of technology and its analysis show the effectiveness of deep learning concepts for gait analysis.