Surveillance systems have been widely deployed in public places, for example to maintain order in a train station with stong people stream, to detect potential dangerous object in airport, to recognize a theft in a store, etc. Traditional way in which the surveillance videos are watched by a man sitting before the monitors is unreliable, low efficient and costly. Ideally, we would like the system to automatically analyze the surveillance videos for reporting the speciall situation.
Objects and person of interest are detected by a deep learning-based method, such as Faster RCNN. Each detected object is labeled with an owner. A background model is utilized to find static objects. If the object's ower disappears from the surveillance scene, an alarm for abandonment is triggered. Further events around the abandoned object are analyzed. If anyone attempts to do anything on the under watched object, the person is verified whether he is the owner. If not, a warning for an un-owner moving the object is triggered. When the person is going to leave the surveillance scene but the missing object is not detected within the scene again, the behavior is recognized as stolen.
In the following sequences, object tracking is performed only with the usage of keypoint features. So the object position which is not suitable in many tracking situations was not used.