Real-time adversarial GAN-based abnormal crowd behavior detection

Detecting abnormal events in the crowd is a challenging problem. Insufficient samples make those traditional model-based methods cannot cope with sophisticated anomaly monitoring. Therefore, we design a real-time generative adversarial network plus an add-on encoder to deal with the continually changing environment. After the generator reconstructs the compressed pattern to generate the design to the latent vector, a discriminator is used to construct better videos by minimizing the adversarial loss function. We calculated the abnormal score by the distance between the two underlying patterns encoded by the first and the second encoders. The unusual event is detected when the anomaly score is above the threshold. To accelerate the processing efficiency, we introduced the grouped pointwise convolution method to decrease the computing complexity. The frame-level and video-level experiments on the benchmark dataset show the accuracy and reliance of our approach. The acceleration approach can increase the efficiency of the network with only limited accuracy loss.