Pedestrian Tracking Research In Automated Driving Scene

With the renaissance of deep learning,artificial intelligence technology has embarked on a new journey.As an important research direction of artificial intelligence,autonomous driving technology is the product of combining computer vision technology with traditional industry.Autonomous driving technology has deeply penetrated into people’s daily life and can bring huge commercial value.It is a common research hotspot in academia and industry.Nowadays,major Internet companies and automobile manufacturers have devoted themselves to the research of autonomous driving technology.As an indispensable part of automatic driving technology,pedestrian tracking plays an important role in improving the safety of automatic driving.Compared with the general detection and tracking,pedestrian in the automatic driving scene occupies a smaller proportion of the picture,the background environment and movement posture are more diverse,the position changes faster,and the real-time requirement is higher.The general detection and tracking algorithms are difficult to meet the pedestrian tracking task in the automatic driving scene.Therefore,this paper studies pedestrian detection and tracking,and designs a pedestrian tracking algorithm suitable for automatic driving scenario.For pedestrian detection,we design a special lightweight pedestrian detection model.In this model,the multi-scale down sampling module proposed in this paper is used to reduce the loss of information in the down sampling process.And a spatial pyramid pooling module is attached the end of the last down sampling to further enrich the feature information and increase the detection performance of small pedestrian.Experiments on the comprehensive data set of BDD 100 k and City Person show mAP of the lightweight pedestrian detection model designed in this paper can reach 64%,and the computation speed on the GeForce GTX 2080 Ti GPU is up to 7 ms/ frame.Meanwhile the model size is only 6MB,and the calculation is as low as 4.47 BFLOPs,which can meet the requirements of automatic driving scenario.For pedestrian tracking,a pedestrian tracking model based on convolutional association network is designed int this paper.This model utilizes the powerful information acquisition ability of convolutional neural network rather than traditional association algorithms to capture the relevance between the targets.And a tracking strategy is designed to deal with the allocation between the detected targets and the existing tracks.Compared with the classical discriminant tracking model,our pedestrian tracking model shares the appearence features with the lightweight pedestrian detection model,which reduces the redundancy calculation and enables the whole detection and tracking process to meet the requirements of the automatic driving scenario.Considering that the position of pedestrian changes dramatically in the automatic driving scenario while relative position between pedestrians remains unchanged,we add spatial relation constraints to the convolutional assotiation network to further improve the pedestrian tracking performance.The pedestrian detection and tracking algorithm proposed in this paper can meet the needs of pedestrian tracking in the automatic driving scenario in terms of speed and accuracy,and can provide some ideas for related research.