Research On Techniques Of The Recognition Of Pedestrians’ Crossing Intention Based On Video Processing

Accurate judgment of pedestrian crossing intention is beneficial for intelligent vehicles to make appropriate driving decisions to reduce traffic accidents.At present,however,there are few studies of the crossing intention of pedestrians.In this paper,the method of the recognition of pedestrian’s crossing intention based on video processing is studied.It is mainly divided into three parts: pedestrian detection,pedestrian tracking and recognition of pedestrians’ crossing intention.The details are as follows:1)Aggregate Channel Feature(ACF)is one of the traditional pedestrian detection algorithms,and it has faster detection speed and better detection accuracy.But it still has high false positive rate.Firstly,considering the depth feature has stronger representation ability than the traditional artificial feature,the classification threshold of classifier used in ACF algorithm is reduced to propose a large number of pedestrian candidate areas.Then,the pedestrian candidate areas are re-evaluated by the trained deep residual network,and the final pedestrian detection results are obtained.Finally,the algorithm is tested in the public database,and the results show that the algorithm can greatly reduce the false positive results of the original ACF algorithm.2)Rapid relative motion between vehicles and pedestrians results in rapid changes in pedestrian scale in images,and short-term occlusion between pedestrians and other objects may occur.Aiming at these two problems,firstly,a background-aware correlation filter.is used to estimate the pedestrians’ position.Secondly,a scale filter is constructed to adapt to the changes of pedestrian scale.Finally,an appearance filter is constructed to evaluate the degree of pedestrian occlusion,and the filters are selectively updated;In this way,a single pedestrian tracking is achieved.Compared with the Kalman filter pedestrian tracking algorithm in public database,the proposed algorithm is more adaptable to the situation of fast scale change and short-term occlusion.3)People can express their intentions through body movements,so pedestrians’ crossing intention can be inferred by observing the changes of pedestrian body parts.Firstly,the pedestrian detector and the pedestrian tracker are integrated into a framework to extract multipedestrian trajectories simultaneously.Secondly,the pedestrian’s body parts in each trajectory are located,and the combination feature of body parts in time sequence is extracted.Finally,a Long Short-Term Memory network is constructed and attention mechanism is integrated into the network to judge whether people cross the street or not.The algorithms is tested in the public database,and the experimental results show that the proposed algorithm has higher accuracy than the benchmark method proposed by the author of the public database.The proposed method of the recognition of pedestrians’ crossing intention can be applied in intelligent driving scenarios to identify pedestrians’ crossing intentions and provide a basis for driving decision-making of intelligent vehicles.