Action Recognition In Complex Background

Human action recognition is a academic valuable and practical valuableresearch area from the perspective of security, surveillance, entertainment orpersonal records. In recent years, the field of action recognition has been developingrapidly. However, the action recognition under complex background did not getenough attention. This paper focus on the research in action recognition undercomplex background.For the deficiencies of the existing action recognition methods, this paperpresents a random sampling method for action recognition. In addition, due to therandom sampling would increase the randomness of recognition, this paper focuseson a further optimization for the random sampling strategy, trying to weaken therandomness of the sampled patch and increase the amount of information carried bythe sampled patch. That is, while taking fewer patches, these patches are morerepresentative of the distinctive characteristics of this type of action.This article uses the following methods:First, present a patch with spatial information and temporal information. Suchpatch not only contains information of the time both before and after it, but also theinformation of the space around it, so the performance of the patch has been greatlyenhanced, for carrying more information only at the cost of increasing limited timecost.Then, use the random sampling method into the action recognition. Use patchtogether with the spatial-temporal patch, it can reducing the test time and does notaffect the recognition accuracy. This paper average accuracy rate by three times toeliminate the chance of random sampling, and take the average value and deviationof the three experimental values and deviations as the accurate value.Experimentally determined the number of samples of each video.Finally, propose an improved strategy for some of the shortcomings ofrandomized algorithms, which is to minimize the randomness. To do this, this paperdefines a new variable to represent the severity of the motion. Through using thisvariable with the heap sort algorithm, the optimized sampling strategy does its best to weaken the randomness of the patch and increases the amount of carriedinformation.This article describes that the rational using of sampling strategy can increasethe accuracy and efficiency of action recognition under complex scenarios, so thatreal-time applications in motion recognition direction has taken a new step.