Vehicle Trajectory Prediction Of Night Unmanned System Based On Object Detection

With the rapid development of artificial intelligence technology in today’s society,in order to alleviate traffic congestion and traffic accidents,people begin to apply artificial intelligence technology in the field of motor vehicle driving.Unmanned vehicles emerge at the historic moment,and the detection and trajectory prediction of unmanned vehicles in front of them is an important link to ensure the safety of unmanned vehicles.Especially in nighttime environments,which are inherently accident-prone.Due to low light intensity at night and decreased visual distance,the detection of the vehicle in front of the unmanned vehicle is quite different from the actual situation,and the trajectory prediction of the vehicle in front depends on accurate vehicle detection results.Therefore,it is of great significance to improve the detection accuracy of unmanned vehicles on vehicles in front of them.At the same time,only with the real-time detection speed can the unmanned vehicle avoid the problem of slow response during driving.Therefore,improving the detection accuracy and speed of the vehicle in front of the unmanned vehicle at night is an important issue that needs to be solved urgently.At present,there are few researches on night vehicle detection,and the main problems are as follows:(1)Light intensity at night is low,visual distance decreases,geometric features of vehicles are not obvious,and detection is difficult.(2)The robustness of the vehicle detection algorithm based on traditional features is poor,and it is difficult to be applied to the complex night environment.(3)There are few public night vehicle data sets and lack of features for algorithm learning.(4)Deepening the network can increase the feature extraction ability of the algorithm and improve the detection accuracy.However,the additional parameters introduced by deepening the network are often ignored,which leads to the decrease of the detection speed.To solve the above problems,this paper focuses on improving the accuracy and speed of night vehicle target detection.Finally,vehicle detection was applied to vehicle trajectory prediction,and the coordinate sequence of vehicle prediction frame was extracted as the data set of trajectory prediction model training by using the vehicle detection results.Specific work is as follows:(1)In terms of the data set of vehicle detection algorithm,in order to improve the practicability of the data set,this paper refers to the safe sight distance of vehicles in the process of driving,and collects night vehicle images within the range of the safe sight distance as the original data set of this paper.Moreover,Retinex image enhancement algorithm was used to reduce the hiding effect of light on the geometric features of vehicles and enhance the contour features of vehicles,so that the target detection algorithm could extract the feature information of vehicles more effectively in the training process.(2)In terms of night vehicle target detection algorithms,four mainstream target detection algorithms including Faster RCNN,SSD,Transformer and YOLOv5 are compared in this paper,and YOLOv5 algorithm is selected as the basic algorithm of this paper.On this basis,by introducing a variety of attention mechanisms into the FPN structure of Neck layer of YOLOv5 algorithm,the MA-FPN structure is designed,which increases the feature extraction ability of the algorithm and improves the detection accuracy of the algorithm.At the same time,deep separable convolution and void convolution are introduced into PAN of Backbone layer and Neck layer of YOLOv5 network.The structures of DS-CBL and D-PAN are designed to balance the additional computational parameters added by introducing attention mechanism,achieve algorithm parameter pruning,and improve the detection speed of the algorithm.Experimental results show that the performance of the proposed algorithm is better than the current mainstream target detection algorithm,and the detection accuracy is improved by 5.2% and the detection speed is improved by 9.1% compared with the original YOLOv5 algorithm.It can provide basic guarantee for subsequent vehicle trajectory prediction.(3)In the aspect of vehicle trajectory prediction,this paper applies vehicle detection to trajectory prediction,and uses the results of vehicle detection to train the trajectory prediction model.Firstly,in order to obtain continuous vehicle trajectory information and make it match with the vehicle time,the tracking algorithm and the object detection algorithm in this paper are combined.The pixel size of the vehicle detection frame was converted to the real size by the monocular vision principle to realize the distance measurement and velocity measurement of the vehicle in front.The coordinate sequence of the vehicle detection frame was extracted as the data set for the training of the trajectory prediction model,and the trajectory of the vehicle in front was predicted under the LSTM time series model.Finally,a good trajectory prediction effect was obtained,and the mean square error of the model was about 0.01.