This paper presents a lane detection and tracking system based on a novel lane feature extraction approach and the Gaussian Sum Particle filter (GSPF). The proposed feature extraction approach is based on the fact that by zooming into the vanishing point of the lanes, the lane markings/boundaries will only move on the same straight lines they are on. Objects other than the lanes in the frame do not share this property and can be ignored during the model parameter estimation. This algorithm is able to iteratively refine various traditional feature maps and to operate with curved roads. The tracking part of the system is initialised by a deformable template matching algorithm. Three types of tracking algorithms are compared in our study: the original Sequential Importance Resampling (SIR) particle filter, the Gaussian Particle Filter (GPF) and the Gaussian Sum Particles Filter (GSPF). The GSPF achieves the best performance by integrating a novel likelihood function and an intuitive parameter selection process. Both the GSPF and GPF provide improved tracking performance and require less computational power than the SIR. It has also been found that the detection and tracking performance is enhanced significantly by incorporating the refined feature map.