Research On Technology Of VFOA Detection In Visual Behavior Feature Fusion

Visual Focus of Attention(VFOA)refers to the direction and focus point of visual attention,which is determined by the combination of head posture and gaze information.Usually,the point of gaze is determined as the position of VFOA.With the rapid development of the artificial intelligence(AI),it is a trend of AI in recent years to build an all-round perception intelligent era.The visual,reasoning and emotional factors are integrated in VFOA detection technology.Therefore,there exist not only deep theoretical value,but also broad application prospects in the research on VFOA detection.Firstly,an overall scheme of VFOA detection system is designed.Aiming at the problem of inaccurate eye location in low-quality images,a method based on prior MTCNN is proposed to locate the two pupils.MTCNN-mxnet,which has good robustness to head pose deflection,illumination change and occlusion,is selected for face detection and five key points regression of human face(Pupils,nose tip,left and right corners of mouth).According to the prior knowledge of MTCNN-pupils,an eye candidate region is segmented.And the gray value and gradient integral projection are carried out on this region.The combination of projection curve extremum coordinates and two points nearest to the left and right pupil positions of MTCNN are taken as the coarse location for pupils.Then the average coordinate positions of the corresponding MTCNN-pupils position are obtained respectively,which are used as the precise location positions of left and right pupils in this paper.Finally,the edge and corner detection algorithms are used to accurately locate the four canthus in the eye region combining with the pupils,which lay a foundation for the establishment of the next gaze estimation model.Experimental results show that the method proposed in this paper improves the average accuracy by1.11% and there is a better real-time performance in low-quality images compared with MTCNN.Secondly,a dynamic Bayesian network model for VFOA detection based on behavior features fusion is established.The hybrid Bayesian network is selected by analyzing and comparing the advantages and disadvantages of other algorithms.Bayesian sub-models for head pose and gaze detection are established respectively.Aiming at the problem of data missing caused by extreme attitude and dynamic scene,a prediction sub-model is added.The weighted fusion of each sub-model improves the recognitionaccuracy and reduces the recognition error.In order to further enhance the adaptability of the VFOA detection algorithm to dynamic attitude changes,the relevant parameters and weight factors of the sub-models are dynamically updated incrementally.Experimental results show that the proposed algorithm based on behavioral feature fusion can effectively estimate the VFOA of experimenters,which is more robust to head deflection and distance change.Finally,the integrated implementation of VFOA detection is completed on a platform of intelligent service robot.The software and hardware parts of the system are designed.The recognition results of VFOA detection are transformed into corresponding control instructions to realize the human-robot interaction.Experimental results show that the VFOA detection method proposed in this paper can effectively control the motions of service robots and has strong real-time performance,reflecting its practical value.