Real-time Detection Of Abnormal Face Occlusion Events Under Video Surveillance

As an important branch of computer vision,face detection technology is currently widely used in important fields such as autonomous services and public safety.And with the further development of cutting-edge technologies such as big data,the performance of face detection algorithms based on neural networks has been greatly improved.However,due to the large amount of input nodes and training data of the neural network,the requirements of the algorithm for computing device performance and operating environment are becoming higher and higher.However,in enterprise applications,on the one hand,due to cost control considerations,the actual configuration of software and hardware cannot meet the requirements of the operating environment.Therefore,the actual performance of some algorithms that require large computing resources is not ideal.In addition,face detection under non-limiting conditions is susceptible to interference from factors such as illumination changes,occlusions,and long distances,resulting in poor actual detection performance.Therefore,the practical process of face detection technology is seriously hindered.In view of this situation,this article is based on the actual application,according to the specific application needs of the enterprise,the research of real-time detection of abnormal face occlusion events based on video surveillance.Its main work includes the following three aspects.(1)Aiming at the problem of dynamic shadow interference in the process of moving target detection,a moving target detection method based on background model combined with texture features is proposed.Since the dynamic shadow will move with the movement of the moving target body,it is easy to be extracted as a foreground target in the actual detection process,resulting in a false foreground target.In order to avoid the interference of dynamic shadows,based on the moving target detection algorithm based on the Code Book background model,this paper makes in-depth research on image texture features.Therefore,a moving object detection algorithm based on Code Book background model and SILTP(Scale Invariant Local Ternary Patterns)texture feature is proposed.Experimental results show that the improved method can not only effectively eliminate the dynamic shadows existing in the detection process,but also further improve the robustness of the Code Book background model to changes in lighting.(2)Aiming at the inherent defects in the HOG(Histogram of Oriented Gradient)feature extraction process,a corresponding optimization scheme was proposed.Compared with other feature descriptors,HOG features have obvious advantages such as geometric invariance and optical invariance.But it is also accompanied by inherent shortcomings such as lengthy descriptor generation process,slow extraction speed,and poor real-time performance.In order to achieve the efficient calculation of HOG features,this article optimizes the two important stages in the HOG feature extraction process.The experimental results show that the processing speed of the optimized algorithm is five times that of the traditional method for the same number of detection windows without significant fluctuations in detection performance.(3)A new dual-threshold training mode is proposed for the problem that the traditional Adaboost(Adaptive boosting)algorithm takes too long to train.Although the Adaboost algorithm based on the cascade structure achieves satisfactory performance in both detection accuracy and speed,the training of the classifier is very time-consuming.In order to speed up the training speed,this paper focuses on the double-threshold training method of the optimal weak classifier and makes corresponding improvements.Experimental results show that the improved method not only has a significant improvement in training time,but also is almost consistent with the traditional method in terms of detection performance.