Counting moving people in crowded scenes is in high demand in videosurveillance applications such as controlling traffic flow, schedulingpublic transportations, counting customers in markets, indexing multimedia archives, and detection of overcrowded places.Among theavailable techniques, trajectory-clustering-based methods, and mapbased methods have shown good performances in counting people indensely crowded scenes.The trajectory-clustering-based approachestry to detect every independent motion by clustering feature-pointson people tracked over time, and the map-based approaches generallysubtract the background, and then map the number of people to somefeatures such as foreground area, texture features or edge count.However, existing approaches still have some limitations and difficulties inproducing accurate results.　　The trajectory-clustering-based approaches fall into trouble in complex scenes, such as with the close proximity of moving people, freelymoving parts of people, and different object size in different locations of the scene.The map-based approaches suffer from inaccurateforeground/background segmentations, erroneous image features, andrequire large amount of training data to capture the wide variationsin crowd distribution.In this thesis, two approaches are developed forcrowd counting.A trajectory-clustering-based approach, combiningvelocity and location-specific spatial clues in trajectories, is proposedto cope with limitations of existing trajectory-clustering-based approaches.Also, a new map-based method, using motion statisticsof feature-points, is proposed to accurately estimate the number ofmoving people in crowds.　　The proposed trajectory-clustering-based approach firstly extracts thevelocities of the trajectories over their life-time.To alleviate confusionaround the boundary regions between close objects, extracted velocity information is utilized to eliminate unreal-world feature-points onobjects bonndaries.Then, a function is introduced to measure thesimilarity of the trajectories integrating both of the spatial and thevelocity clues.This function is employed in the Mean-Shift clusteringprocedure to reduce the effect of freely moving parts of the people.To address the problem of various object sizes in different regions ofthe scene, we suggest a technique to learn the location-specific sizedistribution of objects in different locations of a scene.The experimental results show that our proposed method achieves a good performance.Compared with other trajectory-clustering-based methods,it decreases the counting error rate by about 10％.　　The second proposcd approach, the map-bascd one, utilizes motioninformation of feature-points to estimate the number of moving people in a crowd.Simple feature-points are tracked within the sceneand the amount of moving feature-points is used as a clue to theforeground area, which can provide a coarse estimate of crowd size.Moreover, motion trajectories of feature-points are utilized to capturesome clues about the level of occlusion in the scene.Two statistical features are extracted that are highly correlated with the amount(level) of occlusions in a crowd.Finally, a classifier is trained tomap the extracted features to the number of people.The proposedapproach is evaluated on a large pedestrian dataset.The experimental results and comparisons with existing approaches showy that theproposed method achieves a superior performance.Furthermore, theproposed approach is improved by using local features (local with respect to the groups of people, obtained by using a clustering scheme)instead of holistic features from the scene.This system is shown tobe quite robust and generalizable, as it is capable of extrapolating tocount crowds not encountered during training and can be trained ona small training data.