Real-Time Detection Based on Improved Single Shot MultiBox Detector

In recent years, the convolutional neural networks arc widely used in the field of object detection. However, these methods based on convolutional neural networks require a large amount of calculations, so that it is difficult for these methods to run on platforms with limited computation. A fast object detection method is proposed based on single shot multibox detector (SSD), namely Faster-SSD. The method realizes the real-time detection and high accuracy with limited computation. The basic network of SSD is replaced with ResNet-31. In the stage of generating the prediction frame, first obtain the prior boxes which satisfy the condition, and then generate the prediction frame of the corresponding category. The variable minimum threshold is proposed to reduce the amount of computation. Finally, the online hard example mining is applied to remove the simple samples. Experimental results show that the Faster-SSD gets 14 frame/s on NVIDIA Jetson TX2.