CMNet: A Connect-and-Merge Convolutional Neural Network for Fast Vehicle Detection in Urban Traffic Surveillance

A wide variety of vehicle detection approaches using the deep convolutional neural network (CNN) have achieved great success in recent years. However, the existing CNN-based feature extraction algorithms, especially residual network, cannot obtain powerful semantic information in the vehicle detection task, and thus suffer from the problem of a missing detection, error detection, or repeated detection. In this paper, we present a connect-and-merge convolutional neural network (CMNet) for fast detecting vehicles in complex scenes. First, we propose a connect-and-merge residual network (CMRN) for performing feature extraction. Specifically, the CMRN assembles residual branches in parallel through a connect-and-merge mapping: Connect the input to the outputs of two residual branches separately (Connect), and merge the outputs of the connection as the input of the subsequent residual block (Merge), respectively. Second, we present a multi-scale prediction network (MSPN) to accurately regress the vehicle shape and classify vehicle fine-grained categories. In addition, the feature maps from the CMRN are merged with their corresponding upsampled features from the MSPN using concatenation. Thus, we can improve the information flow of the framework and make it easy to train. The experimental results on the KITTI dataset and the UA-DETRAC dataset demonstrate that the CMNet can achieve efficient detection performance in terms of accuracy and speed for the real-world traffic surveillance data.