Target Detection,tracking And Prediction In Unmanned Driving In Haze Weather

With the rapid development of artificial intelligence,people are increasingly focusing on automatic driving technology,focusing on reducing accidents and economic benefits.Among them,the accurate detection,tracking and trajectory prediction of traffic participants around driverless vehicles,namely vehicles,pedestrians and cyclists,has become an important research direction.However,due to the complex road conditions,uncertain environment and diversity of environmental targets,there is a great deviation between the target detection,tracking and prediction of the above traffic agents and the reality.Therefore,how to overcome the influence of the above problems and realize the detection,tracking and prediction of traffic participants with high precision and high robustness is an urgent problem to be solved in automatic driving.This paper will study the target detection,tracking and prediction of unmanned driving in haze weather.In haze weather,the clarity of the collected pictures is poor,which reduces the accuracy of target detection,and then affects tracking and prediction.Therefore,accurate target detection can still be carried out in the haze environment,which is the basis of follow-up tracking and prediction,and is of great significance to prevent traffic accidents and ensure personnel safety.Moreover,accurately predicting the trajectory of surrounding participants is a necessary prerequisite for safe automatic driving,which can provide decision-making judgment for driverless vehicles.Firstly,the image is de fogged and detected.The existing methods for image de fogging can be divided into enhancement based and restoration based.The enhancement based methods ignore the actual factors leading to the poor quality of foggy image and only focus on the direct effect of enhanced image.The method based on restoration,in which the method based on depth of field has high requirements for equipment;The method based on dark channel is easy to over process the sky region and the image tone is dark.Therefore,this paper proposes a dark channel a priori defogging enhancement algorithm for adaptively segmenting the sky region,which combines the advantages of image enhancement and adaptively enhances the brightness and contrast of the defogged image.After subjective observation and objective analysis,it can improve the problems of over processing the sky area,image darkening and contrast reduction.Through the verification of yolov3 target detection algorithm,the algorithm can significantly improve the accuracy of target detection in foggy environment.Then,aiming at the problem that the detected targets are difficult to match each other in video detection,a fog multi-target tracking and matching model based on Kalman filter and Hungarian algorithm is designed.The detection results are matched with the prediction results of Kalman filter through Hungarian algorithm combined with cascade matching to realize multi-target tracking.In the process of multi-target tracking,judge the fog scene every other period of time.When fog is detected in this frame,the subsequent frames will be defogged first and then input into the target detection network.Otherwise,target detection will be carried out directly to prevent defogging of normal pictures.The algorithm can effectively detect foggy scenes and suppress the problems of tracking loss and missed detection.Finally,a general trajectory prediction model suitable for vehicles,pedestrians and cyclists is designed to predict the future trajectory of each target in a certain period of time according to the historical position of each target tracked.Because the convolutional neural network is good at extracting the local features of the image and finally obtaining the high-order features,it is often used in image recognition.Here,the trajectories of driverless vehicles and surrounding traffic participants combined with the raster image of the surrounding road environment information are sent to the convolutional neural network to extract the features,and its full connection layer and loss function are redesigned for training.Finally,a trajectory prediction model that can predict multiple target categories at the same time is obtained.