On-board Pedestrian And Cyclist Recognition Using Deep Learning Methods

In recent years,the application of Advanced Driver Assistance Systems(ADAS)has improved road safety.However,compared with motor vehicles and their drivers,pedestrians,cyclists and other vulnerable road users have been paid less attention.Effectively recognizing pedestrians and cyclists is the base of protecting them.Most of the existing pedestrian and cyclist recognition methods recognize them separately,which often leads to confused cases between two classes.Besides,limited capability of traditional object recognition models makes it difficult to be applied in complicated road environment.To address these problems,a unified framework for pedestrian and cyclist concurrent recognition is established,and relevant key technologies in the recognizing process are well designed,including multiple-instance object proposal method,deep neural network-based object detection,and multiple object tracking with online target-specific self-learning function.In order to generate pedestrian and cyclist object proposals together,a multiple-instance object proposal method is proposed for the complex and changeable pedestrians and cyclists,which is based on the shared salient region and redundancy strategy.According to common characteristics of pedestrians and cyclists,the proposed method can detect their shared salient regions firstly,and generate multiple bounding box instances around the detected shared salient regions based on redundancy strategy,then filter out some invalid proposals according to the geometric constraints of the camera,and finally,output some candidate regions which can overlap ground truths effectively.To detect pedestrians and cyclists concurrently,a deep neural network model is designed based on the Fast R-CNN object detection framework.In order to solve some commonly encountered failures in pedestrian and cyclist detection,such as false positives or negatives,poor effect for small object,and complicated backgrounds,some specific improvement tricks are deployed,including mining hard negatives,fusing multiple scale feature maps,and integrating multiple object proposals.As a result,the deep neural network model can improve detection performance significantly.To enhance the object tracking capability for pedestrians,cyclists and other non-rigid objects,on the basis of particle filter object tracking framework,a multiple object tracking method with online target-specific self-learning function is proposed.The method can combine detection results of the offline trained detector with online target-specific detectors,and can achieve a long and stable track for multiple-category objects.To verify the effect of the proposed pedestrian and cyclist recognition method,a comprehensive pedestrian and cyclist detection database is created in the urban city of Beijing.Then the proposed method is evaluated in a split test dataset.The evaluation results show that,the proposed pedestrian and cyclist concurrent recognition method can not only detect and track pedestrians and cyclists effectively,but also differentiate them clearly,which will provide a wealth of information for the decision-making of intelligent vehicles.