In-vehicle signal processing plays an increasingly important role in driving behavior and traffic modeling. Maneuvers, influenced by the driver's choice and traffic/road conditions, are useful in understanding variations in driving performance and to help rebuild the intended route. Since different maneuvers are executed in varied lengths of time, having a fixed time window for analysis could either miss part of maneuver or include consecutive maneuvers in it evaluation. This results in reduced accuracies in maneuver analysis. Therefore, with access to continuous real-time in-vehicles signals, a suitable framing strategy should be adopted for maneuver recognition. In this paper, a non-uniform time window analysis is presented. By summarizing the average performance time of each maneuver and developing a bi-gram model which reflects the priori and posteriori probabilities of two consecutive maneuvers, the specific window length is adapted based on the prediction of the next upcoming window in processing long duration driving signals, and maneuver recognition is accomplished for each window based on a SVM classification model. All experiments are performed using naturalistic in-vehicle data derived from the UTDrive corpus, which contains approximately 150 drivers operating an instrumented vehicle through predefined routes in the greater Dallas, Texas area. By applying the non-uniform window strategy, experiments could result with 60% overall accuracy on maneuver recognition and especially helps in classifying stop and left/right turn maneuvers.