Research On Vehicle Distance Estimation Algorithm Based On Monocular Vehicle Camera

In the vehicle assisted driving system,the application of computer vision and image processing technology can effectively guarantee the safety performance of the vehicle.The distance estimation technology for vehicles is the core of the vehicle assisted driving system.Distance estimation based on hardware devices such as radar has high cost and is subject to external influences.The traditional monocular distance estimation algorithm requires accurate calibration of the camera’s internal and external parameters.The camera’s small jitter will cause deviation of the result,so it is not suitable for vehicle distance estimation in real-time situation.Under the support of the key science and technology project(2017H6009,2018H0018)of Fujian Province,this paper analyzes the problems of cumbersome calculation,low detection accuracy and complex camera constraints in the existing vehicle distance estimation algorithm.Then the vehicle distance estimation method for monocular vehicle camera is designed based on a convolutional neural network.In this paper,the real-time vehicle distance estimation is realized by the target detection network architecture.Then,the vehicle distance estimation algorithm based on the instance segmentation network is designed on the target detection network to further realize accurate distance estimation.The research contents of this paper are as follows:Firstly,design of vehicle distance estimation algorithm based on the target detection network in the convolutional neural network structure.In order to eliminate the problems of calibration complexity,low detection accuracy and poor real-time in traditional distance estimation methods,this paper uses Faster R-CNN as the network infrastructure.This paper uses the average vehicle height as a reference,selects the appropriate vehicle dataset,fine-tunes the network model structure and parameter information designs a vehicle distance estimation algorithm and improve the accuracy of the distance estimation for monocular vehicle camera.Experiments show that the method proposed in this paper is 5% lower than the average normalized error of the traditional methods.The average time of distance estimation is less than0.08 s which can satisfy the real-time requirement.Secondly,vehicle distance estimation algorithm based on instance segmentation network for monocular vehicle camera.In the target detection network,the bounding boxes may have a deviation due to the redundant items and different vehicles types.Although the absolute distance of vehicle can be calculated using the average vehicle height as a reference,there is an uncertainty deviation from the ground-truth.Therefore,this paper proposes a vehicle distance estimation method for monocular vehicle camera based on instance segmentation network and retraining the network model for segmenting the vehicle by Cityscapes dataset.The target vehicle filled with a mask is regarded as a foreground,and the remaining items are backgrounds to highlight the required vehicle target information in the traffic scene and eliminate the distance error caused by the detection frame deviation.Combined with the vehicle classification network and based on the output vehicle size of the vehicle classification network,the absolute distance of the vehicle is estimated by combining the mask of the vehicle in the instance segmentation network.Experiments show that the vehicle distance estimation method based on the instance segmentation network is reduced by 2% compared with the method based on the Faster R-CNN network.In summary,this paper proposes vehicle distance estimation methods for the monocular vehicle camera with a convolutional neural network as the framework.The proposed method based on convolutional neural network combines the size information of vehicles with the target detection network architecture and the instance segmentation network architecture.The vehicle distance estimation methods proposed in this paper is not constrained by the calibration of the internal and external parameters of the camera and meets the real-time traffic requirements and has a high accuracy of distance estimation.Therefore,this paper has important research value and practical prospects in ensuring traffic safety.