Object Detection In Traffic Scene Based On Machine Learning

Cars have become widely used mobile vehicles under the impetus of urbanization,bringing convenience to people’s travel.However what comes with it is the increasingly complex traffic environment and the increasingly serious traffic safety issues.Using an effective traffic scene understanding mechanism to endow vehicles with certain intelligence to help the driver avoid safety risks has become a research hotspot to reduce the incidence of road accidents.In the understanding of traffic scenes,the detection of objects is an important research direction.Object detection refers to the location of possible objects in the road traffic environment,and provides more complete scene information.Due to the complicated traffic scene,there are many interference information such as illumination and bad weather effects.It becomes more difficult to rely on manual realization of object detection.Machine learning as a detection method with certain adaptability has become the focus in academia and industry within recent years,and it has been widely utilized.There are two main research methods for object detection in traffic scene based on machine learning: object detection method based on artificial designed features combined with machine learning classifier and object detection method based on auto-extracted feature based on deep learning.Based on the above research methods,this paper proposes object detection methods separately based on machine learning and deep learning,aiming at the problems of detection efficiency and generalization performance in complex urban environment with traffic scene images.The main research contents are summarized as follows:(1)For the object detection method of machine learning,a detection method based on hierarchical support vector machine is proposed.The taillights with obvious background distinction are used as the salient features to determine objects,extracting the characteristics of the taillights and matching the taillights according to the rules,fusing the multi-space effective information and implementing the taillights detection and status recognition by using the hierarchical support vector machine model.We also give experimental verification results.(2)For the object detection method of deep learning,based on the improvement of R-FCN structure,a multi-scale cascade R-FCN detection method is proposed.The semantic information of different levels is merged through skip-layer connection and cascaded strategy in the network.Then the batch normalization is added to accelerate the convergence speed of network.Finally,the improved non-maximum suppression is utilized in inference stage,thus the accurate detection results are obtained and the excellent performance of the model is verified by experiments.(3)Furthermore,considering the computational complexity of the R-FCN structure,a simplified Light Head R-CNN structure is used instead,and a multi-scale cascade light head R-CNN detection framework is proposed.The network convergence strategy is improved.With the computational resource consumption reduced,the detection accuracy is increased.The comparison experiments are carried out on the taillights detection and the multi-object detection in traffic scene,and the results show the accuracy and efficiency,also the superior generalization performance of the detection network.