Research On Pedestrian Behavior Recognition Based On Transfer Learning

In recent years,major internet companies and automobile companies has been focusing on the research on unmanned pilot technology and autopilot system,of which pedestrian behavior recognition is the assurance that unmanned vehicles can drive safely.It is of urgent research significance to improve the precision and the real-time property of pedestrian behavior recognition.This paper mainly studies pedestrian behavior recognition on the basis of transfer learning,and proposes pedestrian behavior recognition algorithm based on multi-source domain transfer learning and lightweight pedestrian behavior recognition algorithm based on deep transfer learning.The main contents of this paper are as follows:(1)Pedestrian behavior recognition algorithm based on multi-source domain transfer learningIn order to solve the issues of insufficient data sets and uneven data distribution,a pedestrian behavior recognition algorithm,based on multi-source domain transfer learning using Wasserstein metric,was proposed.The Wasserstein metric algorithm was applied to increase the compactness of domain distribution,reduce the variation between the feature distribution of source domain and target domain,and improve the generalization performance of the model on target domain.In addition,the multi-source domain self-adaptive regulation mechanism is also integrated.By increasing the amount of regulatory factors,in each iteration of the network,the target domain can take feature information from source domain that are more relevant to use for reference,thus improving the recognition of the target domain.Experimental results show that the pedestrian behavior recognition algorithm proposed in this paper can achieve the precision of 76.6%,which higher than other models.(2)L-MobileNet pedestrian behavior recognition algorithm based on deep transfer learningTo increase the detection speed of the pedestrian behavior recognitions and to tackle the large convolution computation and nerve necrosis issue in MobileNet v3 small algorithm,the algorithm is converted to lightweight version and is integrated into knowledge distillation,in addition,a pedestrian behavior recognition algorithm based on deep transfer learning using LMobileNet small model is also proposed.The new algorithm replaces MobileNet v3 point-bypoint convolution with ghost module to reduce the computation.ELU activation function was used to instead of RELU activation function to avoid the occurrence of partial neuron necrosis in MobileNet v3 model.The knowledge distillation method is utilized to instruct the training process.The multi-source domain adaptive model based on Wasserstein metric is taken as the teacher model while the improved L-MobileNet small model is taken as the student model.Therefore,the knowledge of the multi-source domain adaptive model is transferred to the LMobileNet small model.Thus,the detection speed of pedestrian behavior recognition is increased and the real-time performance of the detection process is enhanced.Experimental results show that the L-MobileNet pedestrian behavior recognition algorithm based on deep transfer learning has the lowest number of parameters,which is 4.1M.The inference time is also shorter than other algorithms,such as ShuffleNet v2,GhostNet and Efficientnet....It is verified that the proposed L-MobileNet pedestrian behavior recognition algorithm based on deep transfer learning has better detection accuracy and faster detection speed comparing with other algorithms.