Research On Video Intelligent Recognition Technology For Home Safety

With the aging of the population becoming more and more serious,the number of elderly people in need of care is increasing.Because young people and carers cannot take care of them all the time,and the elderly are weak and osteoporotic,they are prone to accidental falls and other behaviors.In severe cases,life is threatened.If it is not found in time,the prime time for rescue may be missed.Therefore,in order to effectively solve the care problem of the elderly,there are many detection methods,such as detection systems based on wearable sensors,which have good detection performance,but the elderly are not used to wearing them,and certain false alarms will be caused during exercise.Ambient-based detection systems,although not required to be carried around,are difficult and expensive to install.In order to solve the above problems,a detection system based on video surveillance is proposed.The system is low in price,easy to install,and can better identify the physical condition of the elderly,but the system may leak privacy.Therefore,it is best to judge the fall behavior of the elderly on the end side,avoid uploading it to the background,and eliminate the psychological pressure caused by it.This paper proposes a fall detection system based on the ZYNQ platform,aiming to realize the detection function on the end-side,and to make the system have a certain recognition real-time and accuracy rate.The main research content includes the following three aspects:1.Use the python script to establish the fall detection data set of lmdb type,select the appropriate network training structure,conduct the convolutional neural network model training under the caffe framework,and adjust the parameter values and learning rate until the model with high recognition rate is trained,and test and analyze.After the test is passed,the weight and bias values can be extracted from the model.2.Build a hardware system platform based on ZYNQ according to the overall structure of the system,use vivado software to design and implement modules such as video capture,display,Bayer to RGB,and then build a modular hardware platform.Furthermore,the SDSo C software is used to implement the convolutional neural network and deploy it to the development board,the camera collects images,and the detection and judgment of the fall behavior are completed on the development board.3.Using hardware to implement high-computing-density algorithms can speed up the recognition efficiency.Compare and analyze the system performance of the pure software solution,hardware acceleration and optimization solution,model quantification and convolution algorithm optimization solution,including real-time performance,accuracy,resource utilization,etc.,considering all aspects,and finally choose model quantization and algorithm optimization plan.After testing and analyzing the above three implementation schemes,it is verified that the system basically meets the goals of real-time and accuracy.After accelerated optimization and quantification,the average single-frame recognition time is 1.68 s,and the recognition rate reaches 95.5%.The system can better detect the fall behavior of the elderly,and is suitable for application to elderly activity centers,homes,nursing homes and other areas to better protect the physical safety of the elderly.