Urban mobility data is important for urban analytics and transportation modeling, but real-world trajectory datasets are often limited by privacy constraints, restricted access, and missing or unreliable observations. In this paper, we propose a graph-based abstraction method for generating synthetic pedestrian trajectories from raw GPS data. The approach first clusters trajectory points into spatial regions, then builds a directed weighted mobility network from transitions between clusters. New trajectories are generated through a biased random walk on this network and mapped onto the road network using a routing engine to obtain geographically plausible routes. This abstraction also helps mitigate noise and incompleteness in the original data by reconstructing plausible movement patterns from aggregated spatial structure. We evaluate the method on two trajectory datasets from the Paris area and compare clustering techniques using heatmap-based spatial metrics. Our results show that agglomerative clustering provides a more suitable graph structure than DBSCAN for this generation task. Graph connectivity and compactness, shaped by spatial abstraction, play a key role in the quality of the generated data.