Usage of Asymetric Small Binning to Compute Histogram of Oriented Gradients for Edge Computing Image Sensors

In case of multiple imaging sensors used in different networks (home security, surveillance, automotive, industrial), there is a challenge to perform object detection algorithms in real time, even on the cloud, for a large number of sensors. This is why there is an intensive effort in the industry to move object detection processing on the edge, with the benefits of reducing the bandwidth needs and allowing for scalability in large networks. In this paper we present a hardware friendly optimization technique to compute Histogram of Oriented Gradients (HOG) on the edge, by approximating the HOG orientation as a multitude of small bins. The technique is implemented in RTL for FPGA or ASIC and serves as the first step in a standard object detection algorithm (using Histogram of Oriented Gradients as feature extractor and Support Vector Machine as the detection algorithm). We verified the results of the proposed optimizations for errors by comparison to a reference method and the overall object detection algorithm for robustness.