Moment Based Multi-lane Detection And Tracking

Over the past century,vehicles have brought many conveniences to people.At the same time,manual errors or fatigue driving have caused frequent traffic accidents,which have brought great social pain and huge economic losses.People are looking forward to autonomous driving car.At the same time,as an inevitable trend in the development of the automotive industry and artificial intelligence,autonomous driving car can not only bring huge market benefits to the automotive industry,but also a peak that artificial intelligence must conquer.Therefore,in recent years,more and more self-driving taxis,shuttle buses,and mobile retail vehicles have appeared in common scenes such as urban roads,parks and squares.As the main component of the road,the lane boundary makes the lane boundary detection technology one of the core technologies of the automatic driving system.On some unclearly marked roads,when there are other road markings on the road,or the illumination is insufficient,the existing lane boundary detection algorithm still cannot fully and clearly extract all lane boundaries on the road.However,the above-mentioned situations often occur in the actual environment and pose a great threat to the safety of automatic driving or assisted driving.With the rapid development of neural networks,although some neural networks have shown impressive results in lane boundary detection,the performance of neural network detection results mostly depends on a large number of lane images in different scenarios.Although neural network-based algorithms can accurately detect the lane boundaries in the trained scene,it is difficult to achieve satisfactory results for the untrained scene,which is also the fundamental problem of the neural network.Compared with single-lane boundary detection,by detecting all lane boundaries in front of the vehicle,the exact position of ego-vehicle and surrounding vehicles can be determined.In addition to realizing lane keeping,the intelligent driving system can also realize autonomous lane change and autonomous overtaking,with all lane boundaries detected.This thesis proposes a novel multi-lane detection algorithm to detect all lane boundaries on the road ahead of the vehicle.Different from the common Hough Transform to detect straight boundaries,the proposed method uses geometric moments to extract the centroid of each lane boundary,and calculate the deflection to obtain a more accurate lane boundary.This thesis also proposes a multi-lane tracking algorithm based on geometric moments and Kalman Filter,which realizes the extraction of multi-lane boundaries in subsequent frames with minimal calculation and ultra-high accuracy.In addition,in view of the distribution characteristics of lane boundaries in road image,this thesis proposes a local optimal binarization algorithm based on classic algorithms,which has achieved good binarization effects in a variety of complex environments and has strong practicability.In order to prove the superiority and practicability of the proposed method,it is compared with some start-of-the-art multi-lane detection algorithms based on neural network and non-neural network in detail.The experimental results show that the proposed method has achieved high detection accuracy and detection rate in many challenging road scenes.Finally,the proposed method is applied to a self-developed auto-driving car.The proposed method can quickly and accurately provide the multi-lane information of the road,whether it is day or night,so that the car can drive smoothly and autonomously on road.