The difficulty of counting the number of people using video surveillance lies in the correct segmentation of each person in crowded situations.For this problem，we train a classifier to detect head candidates so as to distinguish between people walking closely under the overhead mounted camera.The static false detected candidates are removed according to the motion feature of the head，while the dynamic ones are eliminated due to their low detected frequency and discontinuous response positions.Finally，an easy and effective crossing tracking scheme is proposed for counting.We conducted experiments in various complex situations.It shows that the counting rate is over 95%.Compared to the traditional method，the proposed means can segment and count the people more accurately in cases where multiple people are crossing or crossing with items.It is therefore much more suitable for people counting in real life situations.