Research On Vehicle And Pedestrian Detection Algorithm Based On Machine Vision

With the development of deep learning,general object detection has made significant progress.However,since the detection of vehicle and pedestrian in autonomous driving distinguishes from the general object detection,there are still many challenges in vehicle and pedestrian detection in driving environment.Vehicles and pedestrians based on video or images captured by road conditions have a wide range of scales.The small and medium-size objects account for a large proportion.The resulting consequence is that the existing object detection algorithms have low detection accuracy for them.Against this background,this research focuses on the problem of multi-scale vehicle and pedestrian detection under real road conditions,and proposes a detection algorithm which adapts to multi-scale object and improves detection accuracy on small and medium-size object.Inspired by the classic general object detection algorithm Faster R-CNN,the present study proposes a multi-scale feature fusion vehicle and pedestrian detection algorithm.Firstly,we analyze the scale distribution of vehicle and pedestrian,and derive the influencing factors of the receptive field in theory.The experiment also verifies the effect of receptive field on detection accuracy.In order to detect different scale objects,because different layers has different receptive fields.we propose a multi-scale feature region proposal network,which provides a region proposal network with wider range of receptive field and generates high quality proposals.Secondly,so far as the problem of insufficient feature on the high-level feature map of small and medium-size object is concerned,we design a feature fusion module which combines the low-level features with details and the high-level features with semantic information to improve the classification accuracy.Finally,considering that the complexity of the actual driving environment leads to many difficult samples,we introduce the hard example mining strategy and use Focal Loss to transform the loss function to realize hard example mining,which further improves the detection accuracy based on multi-scale measures.In order to verify the rationality and effectiveness of the algorithm designed in this thesis,we add a vehicle information to the Caltech dataset based on actual driving environment,and combine some popular training techniques to realize an end-to-end training multi-scale vehicle and pedestrian detection algorithm.Through experiments,the in-depth study of the measures proposed in this thesis can improve the accuracy of multi-scale object detection.The results show that the multi-scale feature fusion vehicle and pedestrian detection algorithm proposed in this thesis has achieved very competitive results,especially for the detection accuracy of small and medium-size object.At the same time,the present research has certain reference significance for the detection of special object with large changes in scales.