Research And Implementation Of Pedestrian Detection For Advanced Driver Assistance System In Complex Scenes

With the continuous improvement of economy,the rapid increase in car ownership has led to a large increase in traffic accidents.As one of the vulnerable groups among traffic participants,pedestrians are always seriously injured or even dead in a traffic accident.As one of the core functions of vehicle advanced driver-assistance systems,pedestrian detection system can effectively reduce the risk of pedestrian injury and is significant to protect pedestrian.Now,pedestrian detection has become one of the hotspots of automotive industry and research institutions.Based on the study of current vision-based pedestrian detection algorithms,this paper proposed a real-time pedestrian detection method for advanced driver-assistance systems in complex scenes,and built a test platform based on NVIDIA Jetson TX2 developer suite,proved the real-time,accuracy and reliability of our method by road test.This paper mainly includes the following aspects:A feature extraction network for pedestrian in complex scenes is designed based on ResNet.A feature extraction network structure is desi gned,batch normalization layer and residual module are added to the network to accelerate network convergence speed and avoid network degradation.A stochastic parametric non-linear activation function is designed,which improved network performance and increased the robustness to data distribution by adding fluctuations in the output activation value.A robust and high performance pedestrian feature extraction network for pedestrian in complex scenes has been created.A multi scale feature fusion pedestrian detection network is designed bas ed on pedestrian feature extraction network.Aiming at the complex and varied background and large change of pedestrian size and posture in complex scenes,pedestrian posit ion is regressed on different scale feature maps respectively.Based on preset pedes trian bounding box regression mechanism,the size of pedestrian bounding box in dataset is aggregated by iterative self organizing data analysis techniques algorithm,and the cluster center are used as the preset pedestrian bounding box size to improve the detection performance of different size pedestrians.A multi task loss function is designed based on cross entropy function and mean square error function,and improved small size pedestrian detection performance by normalizing width and height loss function.Finally an accurate and fast pedestrian detection network has been created.Based on NVIDIA Jetson TX2,a test platform for pedestrian detection algorithm is built and a road test is conducted.In order to meet the real-time pedestrian detection requirement in test platform,single-depth convolution and single point convolution are used to lightweight pedestrian detection network,which improved the detection speed and ensured the detection accuracy.A road pedestrian dataset is established,and it is used to fine tune the network so that the network can adapt to road environment pedestrian detection task;A road test is carried out on the test platform using the trained lightweight pedestrian detection network,test results show that lightweight pedestrian detection network has a good performance in different scenes in road environment,and the detection speed is fast,which can meet the real-time detection requirement of pedestrian detection in advanced driver-assistance systems.