Research On Moving Object Detection And Classification Oriented To Intelligent Monitoring

In recent years, with the development of image detection technology, more surveillance cameras are installed for monitoring the illegal and criminal acts, which can ensure the smooth progress of economic construction and maintain social stability. However, with the traditional manual way to monitor, mass surveillance video will consume a large amount of manpower and material resources, and once the artificial is so fatigue that leave the key video frame out, which will result in serious consequences. Given the shortcomings of the traditional methods, intelligent monitoring systems based on computer vision are increasingly being applied, wherein the motion object detection and classification are important components of intelligent monitoring system. This thesis mainly does research on object detection and object classification, combined with characteristics of the actual situation, we put forward effective overall solutions, and achieved good results. Specific results are as follows:We propose a method of object detection based on moving object detection and binarized normed gradients (BING). Firstly, using the method of gaussian mixture model (GMM) from surveillance video streams, we obtain moving object area; due to the influence of noise, resulting in moving object area containing a lot of noise, and processed by a simple morphological method, we can get a clean moving object area. In order to facilitate subsequent processing (object tracking, object classification, etc.), the output of object detection should contain the smallest rectangle movement of the object area. However, the shadows and other effects cause multiple objects linked or too large rectangular and so on, thus affecting the accuracy of the subsequent processing. The object detection method based on BING can effectively avoid the above drawbacks, which can obtain the minimum rectangle containing the foreground object at a rate of 300 per second. Finally, the method of a combination of GMM and BING, can get valid rectangle.We also propose a moving object classification pipeline based on convolution neural networks (CNNs). Up to now, there has not been a specific database on moving objects, so we established a moving objects database by strict standards, and labeled with motor vehicles, non-motor vehicles and pedestrians, which contains 57,836 training samples and 4,827 test samples. For the resolution of moving objects image is low, we design three layers and five layers of convolution neural network. Since the background changes often, which cause moving objects are not clear, we use histogram equalization for image pre-processing, mAP (average recognition rate) can be increased by 1% in the training set. In China, each category is varied, especially non-motor vehicles, we use unbalanced sample for training, which make the mAP increased by 8% highest. Finally, by multi-model, we fuse two models and obtain a higher recognition rate, making the mAP up to 96% with three-layer convolution neural network.