Real-time object detection in dynamic scenes using recurrent convolutional neural networks

With the growing demand for real-time perception of environment and objects in applications such as unmanned driving and intelligent surveillance, how to realize high-precision and high-efficiency object detection in dynamic scenes has become a significant challenge in the computer vision field. This paper proposes a part-based crowd detection framework hierarchically based on the traditional convolutional neural network object detection framework, which integrates with a multi-level spatiotemporal information fusion module and a feature selection strategy of attention mechanism, building a complex recurrent convolutional network structure. Existing works use the multiscale convolution operation to extract (fine-grained) spatial features in static images. Subsequently, comparing with the original convolutional neural network (CNN) framework, the bidirectional recurrent unit ConvGRU was applied to associate the features of adjacent frames, learning richer time series context information. In addition, the attention sharing module is added to enhance the key area of each frame while suppressing the distracting background, at the same time, the multi-task loss function is used in the detection stage to learn bounding box regression and target classification, so as to make the model have a fast inference speed while ensuring high accuracy. Experimental results show that the cyclic convolutional neural network proposed in this paper achieves higher accuracy and faster inference efficiency on multiple dynamic scene dataset than the mainstream detection models, and has the potential for real time object detection.