Long-distance Multi-target Detection And Location System Based On DM8127

With the rapid development of computer vision, pattern recognition, image processing and artificial intelligence, as well as intelligent network camera more powerful, object detection and location technology increasingly mature, and the practicality of intelligent monitoring has been more and more widely, and monitoring technology will be farther, faster and more accurate. Because distant monitoring not only covers a wide area and detects multi-target, but also has great hidden, it has been widely applied to the field of forest fire prevention, intelligent transportation, and security warning. This thesis aims at detecting and locating distant multiple pedestrians, and implements a real-time and high-accuracy system which is based on DM8127 smart camera platform. The main contents are as follows:1. A pedestrian detection algorithm based on ROI(region of interest) is proposed, which can overcome the shortcoming that existing pedestrian detection algorithm based on the whole picture is too slow. This new algorithm is very fast because distant pedestrians are so small that the areas of ROI are few. It is in favor of real-time detection that moving the frame-difference-based ROI extraction algorithm to the DSP of DM8127 camera and using optimization tools such as EDMA, C64 x instruction set and Cache-based Ping-Pong operation on DSP side.2. Study the pedestrian detection algorithm and find out the HOG_LBP feature is more accurate than the HOG feature in the pedestrian detection. Comparing with the HOG_LBP/HOG methods based on the whole picture on accuracy and time, the HOG_LBP method based on ROI is better in the pedestrian detection.3. For locating multi-target, reference-points-based single-camera locating method can locate multi-target but is not accurate, however multi-camera locating method is accurate but can't locate multi-target. A new multi-camera locating method based on using reference-points-based single-camera locating method to match targets is proposed, which not only can locate multi-target but also is accurate.4. Study the architecture of the whole system, and add the tracking module. Using Kalman filter helps extract ROI and reduces the undetected possibility. Because the blocked target cannot be detected, using Kalman filter can make up the target and help get a more continuous locating result. Tracking associating targets between frames makes multi-target locating more accurate overcoming bias from space locating.