Surveillance is ubiquitous in our communities which can be utilized to deal with multiple security issues. However, most of surveillance systems still are not intelligent which are mainly relying on security staff’s human labor. Thus, human behavior analysis based on computer vision could tremendously reduce security staff’s workload. To analyze and understand human behaviors in surveillance, the start point is to extract computable features from captured videos based on detected human body, the ultimate goal is to finally recognize human behaviors from motion and event analysis. This thesis presents comprehensive and in-depth empirical approaches for event recognition in surveillance based on distinct Feature Extraction Techniques (FET), namely: Histograms of Oriented Gradients (HOG), Local Binary Pattern (LBP) and Scale Invariant Local Ternary Pattern (SILTP). Each of the FETs is based on local feature descriptor which depends on adjusting the cell size of the ROI to achieve better performance. In this thesis, we find the cell size will influence not only the computational time, but also the precision rate. This thesis utilizes the well-known Weizmann video datasets. While both LBP and SILTP features work very well, HOG has shown its superior performance for human behaviour analytics with five selected events (Walking, Running, Skipping, Jumping and Jacking). The simulated results of three classifiers from WEKA (MLP, k-NN, decision tree) have reflected rightness of the extracted features. In this thesis, the empirical approaches for human behaviour analytics in surveillance reduce human labor tremendously. The contributions of this thesis are: (1) The distinct FET makes the best precision of overall human behaviour recognition at the rate above 97.7%. (2) By adjusting the cell size of ROI, the proposed approaches are able to be accelerated, furthermore, computational time could be reduced.