Aiming at the problem that precision of multi-scale pedestrian detection algorithm is not high

a multiscale pedestrian detection algorithm based on cascade convolutional neural network is proposed. The rectangular convolution kernel is used to extract pedestrian features

and the ratio of width to height of proposals are designed according to pedestrian contour features and the self-adaptive loss function is proposed

which makes the network focus on hard samples and effectively alleviate adverse factors brought by the long tail effect in training network and improves the network generalization ability. Compared with Adapt Faster RCNN

the miss rate of the proposed algorithm declines by 1. 36 % in large-scale pedestrians of Caltech data set and that decreases by 8. 82 % in small and medium-sized.