MINI-SSD: A Fast Object Detection Framework in Autonomous Driving

Abstract

In this paper, a python-implemented infrastructure of a CNN-based multi-object detector in autonomous driving using the single shot detector (SSD) is presented. The infrastructure consists of both training and inference for object detection. The main contribution of this paper is the design of the default anchor boxes tiling that reduce the amount of computations by simplifying the software implementation of the SSD object detector. This simplification is done by reducing the data path of the proposed detector. Moreover, a decrease in the inference time of the detector is the result of using tiled defaults boxes and a small number of layers in the VGG CNN. In addition, the CNN model presents an advantage in terms of high confidence boxes prediction. The proposed approach is faster due to the reduced number of layers and computations. The segmentation design of the input image anchor boxes is introduced to explain the software implementation. In addition, both the training and validation loss variations along the period of the training are illustrated.