Debris Recognition Of Print Circuit Board Images

Explosion crime has become one of the greatest threat to public safety in recent years, especially the way of remote control/timing off-site smart starting.Recognition of electronic circuit boards, which are core components of such devices, provides important evidence to solve such explosion cases. At the same time, in recent years, intelligent video surveillance plays an irreplaceable role in the field of public safety, in which the key technologies include target detection, tracking and event detection. This thesis mainly studies segmentation and recognition of print circuit board images and group event detection in surveillance video, the main achievements are as follows:We propose an object segmentation method of print circuit board images based on probabilistic graphic model and color model of probability sampling.Considering the characteristics of the circuit board itself, we adopt hierarchical segmentation strategy to segment components of circuit board. In the whole level, we find the mainboard candidate pixels with K-means clustering method,and initialize the Gaussian Mixture Model of foreground (components and character object) and background (mainboard), and then exploit Grab cut algorithm for iterative segmentation. In the components level, based on Grab cut segmentation results, we obtain the color models of foreground (component itself) and background (main board and surrounding components) by the uniform sampling method, then use the idea of nearest neighbor to classify pixels and fine correct the segmentation result. The experiments show that our method works well in both the public data sets of circuit boards and the data sets we have collected, which is robust to the shadow and noise in images.To solve problem of debris image recognition of print circuit boards, we improve one matching method of key points based on SIFT features geometric similarity. Considering the characteristics of solid triangle structure, key points are matched when they meet the space constraint of similar triangles in this method, thus mistaken matches of feature points are effectively reduced. We improve this method in PCB image recognition task by modifying the sifting strategy of key point matches and by increasing the strict space constraints.After our improvement, algorithm accuracy is greater than 99% in the experimental tests.We also improve one multi-target tracking algorithm by fitting tracklets and trajectories with Gaussian process regression. We implement a new trajectory analysis algorithm for the detection of group events. In this algorithm,we extract speed, direction and distance of trajectories based on the tracking,and analyze the relationship of those trajectories with time intersection, then judge whether there is a PeopleSplitUp or PeopleMeet event on the rules. In the evaluation of TRECVID-SED 2015, the algorithm is used to detect PeopleSplitUp and PeopleMeet event and achieves the second and the forth rank respectively.