Pedestrian Trajectory Prediction And Head-bowing Abnormal Behavior Detection Based On Deep Learning

At present,more and more vehicles are convenient for people's lives while also making the number of traffic accidents high.Therefore,real-time perception of pedestrians around the vehicle and accurate prediction of their future movement trajectory are of great significance to the drivers,assisted driving,and automatic driving.To this end,our paper designs a pedestrian trajectory prediction method based on the first perspective and an abnormal behavior detection algorithm based on human joint points.With the help of our algorithms,vehicles can drive more safely on the road by avoiding pedestrians more accurately.The main research contents of our paper are as follows:(1)Our paper proposes a pedestrian trajectory prediction algorithm based on the first perspective.Firstly,according to the first perspective scene characteristics,we design algorithms to extract the required features: pedestrian historical trajectory features,scene features,and vehicle self-motion features to help the vehicles better perceive the surrounding environment.Secondly,we encode the above features into visual feature tensors of the same dimension.In the trajectory generator,in order to learn a variety of features,an attention mechanism is introduced into the encoded visual tensor to avoid forgetting defects of the time series algorithm,and then to establish different attentions of feature vectors at different time steps.Finally,according to the state of the decoder at each moment,that is,the feature vector with attention mechanism,the LSTM decoder is used to directly predict the future trajectory of the target person.We verify the effectiveness of the algorithm in our paper by testing on public datasets.(2)In order to detect the abnormal behavior of bowing,our paper proposes a fast and effective algorithm for detecting abnormal behavior of bowing based on human joints.Firstly,we propose a novel method to construct a synthetic dataset,which solves the problem of lacking dataset and corresponding labels.We leverage human joint points by adjusting the coordinates of the left hand and right hand to simulate the posture of deviceholding.Secondly,we make full use of the arm and head information to classify and recognize the coordinates of the above joint points,and then achieve an efficient pedestrian abnormal behavior detection algorithm in complex environments.Finally,the experiment verifies the effectiveness of our synthesized dataset and the accuracy of our algorithm for detecting abnormal behavior in real scenes.