NNDcn-Neural Network Based Deep Crowd Network for Crowd Count

Crowd count refers detecting and estimating the arbitrarily sized human heads in clearly visible sparse and highly cluttered densely populated scenes. Computer vision based deep networks such as convolutional neural network are used in state-of-the-art approaches for counting individuals in highly congested images. Convolutional neural network involves the convolve operation that depends upon the filter size. We proposed a deep convolutional operation neural network based architecture that integrates the information obtained from various receptive fields by using different filter size in the conv operation. In this way, model learns the importance of each extracted features from different locations within image, thus improving the accuracy. Our method uses VGG16 network as frontend and a convolutional neural network having different filter size with dilation rate as backend network to extract rich features form congested scene image. Using this architecture we can generate high quality density map, integrating over which results in count estimation. The proposed model was experimented with state of art ShanghaiTech dataset and able to estimate count in highly congested scenes with higher accuracy.