During the last decades, people detection has received great attention in computer vision and pattern recognition because of its various applications. Though there are thousands of papers provide approaches for people detection, most of them focus on datasets from side view. People detection from overhead cameras, as an important situation for surveillance, is essential to research. This work based on annotated videos from MatchNMingle dataset, which are taken by overhead cameras. The videos were taken in a mingle after a speed-dating event. People are annotated by bounding boxes contain their body. In this crowded social scene, people occlusion is one of the challenge barriers for detection. In this work, we study the relation between people occlusion and detecting performance by experiments. Based on a deep network consisting of GoogLeNet and Overfeat, we analyze the performance of detectors trained with various occlusion distribution at different occlusion level. On the ground of experiment results, we attempt to promote the performance by selection of training data. Apart from this, as an attempt for promotion, we train head detectors by newly collected head annotation. The performance evaluation of these two methods indicates their potential for people detection in crowded scene.