General Target Detection Method Based on Improved SSD

In order to improve the accuracy of general target detection by the surveillance camera in the monitoring area, an improved SSD model named DC-SSD that integrates the deformable convolution network was proposed. This model is based on the SSD convolutional neural network model, and the idea of deformable convolution is used for reference. The deformable convolution module is added into the three high level convolution modules of SSD to make the sampling point position of the high level convolution kernel adaptively change according to the image content, so as to adapt to the geometric deformation of different objects such as shape and size. In the open source neural network framework Caffe, DC-SSD was tested with PASCAL VOC 2007 (train+val) as the training set and PASCAL VOC 2007 (test) as the test set. Under the same training and test conditions, the mAP of DC-SSD reached 72.0%, 4.0% higher than the original SSD model, and the detection frame rate reached 39fps. Experimental results show that DC-SSD model can effectively improve the accuracy of universal target detection and meet the real-time requirements.