Pedestrian Detection by Fusing 3D Points and Color Images

Security Based

In this project, a fusing approach of a 3D sensor and a camera are used to improve the reliability of pedestrian detection. The proposed pedestrian detecting system adopts DBSCAN to cluster 3D points and projects the candidate clusters onto images as region of interest (ROI). Those ROIs are detected by HOG (histograms of oriented gradients) pedestrian detector. Because the DBSCAN groups together 3D points and rejects outlier points correctly, the proposed system has a low false detection rate. The performance is also improved since the proposed system only detects the ROI instead of the whole color image. Recently, many approaches of pedestrian detection are proposed because the technology is widely used in many applications, e.g., surveillance system, door control system, driving assistant system and home care system. The pedestrian detection systems usually use variant perceptions which may be single sensor or multisensor systems. Generally, the acquired information from those sensors can be categorized as 2D information and 3D information. Some existing detection algorithms use only single type information to recognize pedestrians, e.g., camera. The most advantage of the color image based pedestrian detection is cheap because only a camera is needed. But, the main drawback is too false alarms caused by shadows or occlusion because of lacking depth information. Unlike the color image based pedestrian detection, the 3D information based pedestrian detection systems have more accuracy information which can be used to separate objects more exactly. But, recognition of high dimensional features takes more computation time. Furthermore, a Lidar is more expensive than a camera.

This work is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.