Spatial-temporal Deep Learning Algorithms For Dynamic Scene Understanding

Dynamic scene understanding is a subset of computer vision and machine learning.It has been a research hotspot in recent decades.In this paper,an algorithm based on rules for specific events detection in dynamic surveillance scenes is proposed.In view of its existing problems,a spatial-temporal deep learning algorithm for dynamic surveillance scenes is proposed.Then we apply the deep learning algorithm to dynamic traffic scenes and achive anthropomorphic decisions about steering wheel operation for autonomous vehicles.Aiming at the detection of specific events in dynamic scene comprehension,this paper proposes a dynamic scene recognition algorithm based on rule.By analyzing the characteristics of specific events,a specific algorithm is proposed for each different events.Using the classical computer vision algorithms such as optical flow and background modeling algorithm combined with the constraint test based on experience,the corresponding events detection is realized.When the algorithm is applied in the anomaly detection of the crowd in the surveillance scenes,the F-measure of the population anomaly detection is 90.9%,and the robustness to the change of the light is shown in the crowd escape detection task,F-Measure still achieves 61.24% in scenes where the other non-learning algorithms are barely detectable.In order to solve the shortcomings of the above rules based method,this paper proposes an deep learning algorithm for surveillance dynamic scene understanding which is unified for differet events.In this paper,we use the multi-channel three-dimensional convolutional network to extract the high-level features in the dynamic scenes data,and combine these high-level features to classify the contents of the dynamic scene data.And then the events in the surveillance dynamic scene can be detected effectively by combing the network and the experiencial constraints.When the training network,this paper uses the pre-training and fine-tuning way to solve the lack of training samples problem.In the fine-tuning and detection,the use of spatial-temporal block strategy to enhance the detection results.This unified algorithm achieves slightly better performance than the method of setting specific rules for specific events in the anomaly detection of population in the surveillance scenes.In this paper,the above-mentioned spatial-temporal convolutional neural network is applied to the task of understanding the vehicle dynamic scenes for the control volume decision.This paper improves the spatial-temporal convolutional model by appling some improved method for convolutional neural networks.The improved network can be called as spatial-temporal driving decision network.By learning the featues which can help the network understand the characteristics of the vehicle dynamic scenes,the improved network successfully predicted the steering wheel angle like a the experirnced driver.The resulting absolute error in the spatial-temporal driving prediction network is reduced by 0.762 degree relative to the popular two-dimensional convolutional neural network.