نوع مقاله : مقاله پژوهشی
نویسندگان
1 گروه فناوری اطلاعات و ارتباطات، دانشگاه جامع علوم انتظامی امین، تهران، ایران
2 گروه مهندسی مکاترونیک، دانشکده سامانه های هوشمند، دانشکدگان علوم و فناوری‌های میان رشته‌ای،دانشگاه تهران، تهران،تهران، ایران
3 کارشناسی ارشد مهندسی کامپیوتر، دانشگاه جامع علوم انتظامی امین
چکیده
کلیدواژه‌ها
موضوعات
عنوان مقاله [English]
نویسندگان [English]
The existence of security is mandatory in any society and it is the basis for the progress and development of a country as easily and quickly as possible, so all countries try to establish stable security by controlling the level of violence and strife in the society. On the other hand, due to the limitation of manpower, it is not possible to carry out the entire process of providing security through the traditional and common methods of the past, and in this regard, new and up-to-date equipment and technologies must be used. and advanced countries of the world, the use of closed-circuit and surveillance cameras in public places is in this research, an expert system based on two sets of neural network ResNet101 and memory LSTM with the aim of reducing the amount of computation while maintaining proper accuracy, ResNet101 network is presented With a total of 347 layers and through the transfer learning method, it extracts the spatio-temporal features of consecutive video frames, and then the LSTM network with a total of 9 layers is responsible for detecting violent behavior in the video. These two sets have been optimized in terms of the type of layer arrangement, the way of connection and the number of cells in each layer so that they can have the best performance in all video conditions, including low quality, presence of noise and short video, etc. As a result of this intelligent system, they can detect violent behavior in closed-circuit cameras with an accuracy of 86.28% in real-time and instantaneously in low-quality video images of 224x3x224 pixels, and in case of violence, report it to the relevant people. inform In the end, it should be mentioned that the designed system, by reducing the amount of computing while maintaining the accuracy, has been able to perform effective and appropriate online monitoring of low-quality surveillance cameras by using only 22 frames per 5 seconds of video.
کلیدواژه‌ها [English]
)