Object Detection Based on Multi-Channel Deep CNN

Object detection under complex background is a very challenge problems, because it might be very difficult to discriminate target and its background even for human eyes. Thanks to the application of deep learning to object detection, the detection performances have improved a lot year by year. In the paper, we propose an object detection framework based on image fusion using visible image, infrared image and motion image to form three-channel input image. Furthermore, we build a 53-layers neural network by using fused image as input. Cross-domain transfer learning technique is used to train the network on large-scale IMAGENET datasets firstly, then the network is fine-tuned on the small-scale collected image datasets. Both quantitatively and qualitatively experiments are conducted to demonstrate the robustness of our method while maintaining real-time performance.