DETECTION OF CAMERA ANGLE SHIFT IN VIDEO SURVEILLANCE SYSTEMS BASED ON THE REFERENCE POINT METHOD

The article is devoted to solving the problem of detecting camera angle shift in video surveillance systems. Camera shift detection is considered as the problem of determining changes between frames in a video stream. One of the effective approaches to solving this problem is frame matching based on the reference point method (searching for and matching prominent areas in an image). Most scientific papers describe matching static frames or object tracking. However, in urban environments, dynamic frames with moving pedestrians or vehicles often occur. This required additional research into the applicability of the reference point method to these conditions. The study included the implementation of three stages: static camera position (analysis of frames from a stationary camera during the day and night), horizontal camera shift (analysis of frames with controlled horizontal shift from 10° to 50°) and vertical camera shift (analysis of frames with controlled vertical shift from 10° to 50°). The following reference point matching algorithms were selected for studying the method: SIFT, SURF, AKAZE and ORB. To evaluate their efficiency, the following metrics were calculated: the number of matched points (Matches), execution time in seconds (Time), average shift (Shift), proportion of moving points (Shift Ratio), and shift consistency (Shift Consistency). When testing the algorithms on static frames, the average shift of points remained minimal (less than 1 pixel) both during the day and at night. No false positive results indicating camera shift were observed. With vertical camera shift, the average shift of points visually increased, which served as a sign of camera movement. At shift angles greater than 20 degrees, the number of matched points dropped sharply. With a static camera position, the average number of points was stable. In general, the study showed that the ORB algorithm has the lowest execution time (Time), but also low accuracy in calculating the average shift (Shift). The SURF algorithm turned out to be the slowest, although it showed satisfactory results in accuracy. The SIFT algorithm showed good results when working on static frames and vertical displacement, but failed to cope with horizontal displacement. AKAZE showed the most stable results in all tests. Thus, the reference point method can be effectively used to solve the task in video surveillance systems.