Computer vision based surveillance systems have become increasingly important to society. This thesis presents a new approach for computer vision based alarm making systems which detect abnormal events in fixed camera circumstances. The approach contains four functions: namely (1) detecting, (2) tracking, (3) recognizing and (4) alarming. In line with these functions, based on the results of detecting, tracking and recognition, the system will be able to generate alarms automatically. Through the experiments, the related methods and algorithms applied to the proposed approach provide better performance for the purpose of alarm making, thus it could be helpful in reducing the manual labor of security staff. The contributions of this thesis are: Firstly, the shortcomings and deficiencies of the traditional surveillance and alarm systems have been studied. Secondly, computer vision techniques have been utilized to allow the system to work with different environments. Thirdly, dual artificial neural networks have been innovatively deployed for abnormal events detection and to improve the accuracy of alarming to reduce false alarms. The overall result for the false alarm rate of the system developed in this project is 13.8% which is lower that the mainstream 15.27% and also helpful for the management of traffic environments. In future, the improvement of the system will be the working direction for the researcher such as using more training datasets to make the abnormal events alarming system more efficient in terms of abnormal event detection and reduction of the false alarm rate.