| Currently,China is in the stage of an aging population,and the fall behavior of the elderly is one of the important factors affecting their normal lives.Through widely deployed monitoring devices in daily life,computer vision fall detection can effectively reduce the harm that fall behavior brings to the lives of elderly people.Traditional fall detection algorithms based on computer vision work well in simple environments,but they are difficult to apply in complex environments and when pedestrians have obstructions.Aiming at the above issues,this thesis proposes a complex scene monitoring target behavior recognition detection and tracking algorithm in the detection and tracking stage of pedestrian falls,In the fall behavior judgement stage,a VGG-Bi NLSTM target fall behavior detection network in complex scenes was supported.Additionally,a detection system suitable for the fall detection algorithm in this thesis was developed.The main work of this thesis is as follows:(1)A target detection and tracking algorithm for complex scenes with pedestrian occlusion is proposed.By replacing the C3 module in the YOLOv5 backbone network with the CBAM module,and improving NMS to Fuse Soft-NMS,more pedestrian feature and location information is obtained.A feature matching target tracking algorithm based on pedestrian occlusion achieves real-time tracking of occluded pedestrians.By building a complex scene fall detection dataset,the target detection and tracking network are retrained to improve the detection performance.Comparative experiments demonstrate that the network model proposed in this thesis has a better detection rate for pedestrian targets.(2)Proposed a VGG-Bi NLSTM attention-based model for fall detection.By embedding storage units in the LSTM module to generate NLSTM,the robustness of the original LSTM neural network structure was improved,and historical information can be restored and processed for longer periods of time.The improved Bi NLSTM attention network can process temporal information between consecutive video frames,and the target behavior can be prejudged by combining the set feature threshold,effectively reducing the number of feature parameters sent to the Bi NLSTM network,improving network detection efficiency and frame rate.The fall detection model achieved good experimental results on both customized and public datasets.(3)A system for detecting target fall behaviors in complex scenes has been developed.Based on the functional requirements and development direction of the system,Py Qt5 was determined as the front-end development tool and Spring Boot was used for the system design in the back-end.The system has realized real-time and offline fall detection functions,verifying its stability and reliability.Compared with advanced existing algorithms,the algorithm proposed in this thesis has significantly improved target recognition accuracy and the recognition of occluded pedestrians,and further improves the efficiency of fall behavior detection.Additionally,the developed fall detection system is convenient to use and has strong portability,and can be applied to the detection of human fall behavior in complex scenes such as nursing homes,and has strong practical value. |
