ENHANCED VEHICLE DETECTION APPROACH USING DEEP CONVOLUTIONAL NEURAL NETWORKS

摘要

Vehicle detection plays an important role in autonomous driving systems. Recently, vehicle detection methods achieved large successes with the fast development of deep convolutional neural network (CNN). However, due to small size, heavy occlusion or truncation of vehicle in an image, recent CNN detectors still show a limited performance. This paper presents an improved framework based on deep CNN for vehicle detection. Firstly, deconvolutional modules are added at multiple output layers of the reduced VGG 16 architecture to enhance additional context information which is helpful to improve the detection accuracy. Secondly, region proposal modules are applied at different feature maps to address the vehicle occlusion challenge. Due to heavy object occlusion in test dataset, soft Non-Maximum Suppression (NMS) algorithm is used to solve the issue of duplicate proposals. Finally, a deep CNN-based classifier including a region of interest (ROI) pooling layer and a fully connected (FC) layer is used for classification and bounding box regression. The proposed method is evaluated on the KITTI vehicle dataset. Experimental results show that the proposed method achieves better performance compared to other the state-of-the-art approaches in vehicle detection.