Research On Multi-scale And Multi-person Falling Detection Methods Based On Deep Learning

With the expansion of urban scale,the probability of emergencies such as congestion and trampling in public areas also increases.Among a variety of abnormal group behaviors,the group panic behavior caused by falls often lead to extremely serious follow-up consequences.In addition,the fatality rate of falls itself is also extremely high.In this paper,we focus on the fall detection under the important multi-person scenario in group behavior.The existing fall detection algorithms based on computer vision have the following two problems.The first problem is that a complex scene with multiple people cause great interference to fall detection,resulting in low real-time performance.Existing methods mostly focus on fall detection in single-person with simple scenarios.In a scene with multiple people,the information passed into the fall detection network without filtering,will lead to network load overflow and poor real-time performance.The second problem is that the scale of existing behavior recognition network is single,and it is not good for small target detection in multi person complex scene.In the real scene,the range of the surveillance camera is wide,and the remote target accounts for a small proportion in the image.The detailed information of such target is easy to be lost with the deepening of the network,resulting in a high rate of missed detection.Aiming at the problem of high interference detection in multi-person environment,this paper proposes a fall pre-judgment framework based on skeleton information,which filters the non-fall information in the video to reduce the interference of complex environment and improves the real-time performance of the algorithm.First,the human body in the scene is located,and the human body region is introduced into the skeleton point detection network to extract the key skeleton points.According to the subsequent extracted features,whether there is a tendency to fall is judged.In addition,in order to reduce the amount of transmission,the people who may fall were automatically intercepted and retained,and the rest of the information is discarded to improve the real-time detection.In order to meet the requirement of real-time in multi-person situation,this paper proposes a simple and fast fall pre-judgment feature based on skeleton points for the above-mentioned pre-judgment method,which is composed of human neck descent speed and angle changes.Comprehensively judge whether the target has a tendency to fall from these multiple dimensions,this paper selects the potential people who may fall through the given judgement threshold.Through qualitative and quantitative experiments prove that the pre-judgment features and overall pre-judgment methods proposed in this paper are effective and accurate.Aiming at the second problem of small targets missing detection and low accuracy,this paper proposes a multi-scale two-stream heterogeneous convolutional neural network(MSTSN).This paper improves the feature multi-scale fusion of the spatial stream network in the dual-stream network,and fuses the feature information of different scales obtained by convolution of the shallow network and the deep network to obtain more discriminative features and improve the network model in feature extraction.It improves the problem of small target detail loss caused by the deepening of the network layer in the feature extraction process of the network model,and improves the performance of the initial features.Through experimental verification,the improved network can ensure the detection of multi-scale information and reduce the missed detection rate of small targets in complex scenes.At the same time,in order to extract discriminative features to help the network understand the video information,the MSTSN network proposed in this paper uses a dual-stream heterogeneous method to extract spatial and temporal information.The spatial stream basic network uses ResNet to learn the spatial information and appearance of the video,and the temporal stream basic network uses BN Inception to learn movement changes and temporal information.After the two branch networks respectively complete the extraction of spatial information and timing information,the features are merged in MSTSN and then classified to improve the recognition rate of the overall network.In this paper,the above algorithms are tested comprehensively.The experimental results show that the proposed algorithm based on skeleton information in multi-scale and multi-person fall detection can overcome the problems of semantic information loss,low detection accuracy and low real-time response.By comparing with other fall detection algorithms,it is proved that this method can detect the fall behavior quickly and effectively.For the area of multi person scene,the algorithm can timely monitor and prevent the expansion of group events,which has important social value.