SqueezeMap: Fast Pedestrian Detection on a Low-Power Automotive Processor Using Efficient Convolutional Neural Networks

Rytis Verbickas, Robert Laganiere, Daniel Laroche, Changyun Zhu, Xiaoyin Xu, Ali Ors; Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2017, pp. 146-154

Pedestrian detection for autonomous driving is a challenging task that requires careful trade-off between accuracy, storage, computation and energy requirements. In our work, we extend the recent SqueezeNet architecture to pedestrian detection. We show how this network can be modified to obtain detection performance on the Caltech USA pedestrian dataset that is comparable in overall log-average miss rate to other competing models while easily running at 30FPS on an automotive processor with a model size of 3.24MB and within a power envelope of 2W. The extension relies on the observation that precise knowledge of bounding box corners is not necessary to know the location of pedestrians and their approximate size. Rather, a coarse grid based localization is proposed here and acts as a kind of heatmap of pedestrian locations, relying on only one forward pass through the network. The number of new free parameters introduced is small relative to the original SqueezeNet model.