Energy-Efficient Pedestrian Detection System: Exploiting Statistical Error Compensation for Lossy Memory Data Compression

Pedestrian detection represents an important application for embedded vision systems. Focusing on the most energy constrained implementations, systems have typically employed histogram of oriented gradients features and support vector machine classification, which leads to low detection accuracy (a log-average miss rate of 68% on the Caltech Pedestrian dataset). Additionally, single-scale detection is often adopted in these systems for real-time processing, which further deteriorates the detection performance. In this paper, we propose a hardware accelerator achieving substantially higher detection accuracy by employing aggregated channel features (ACFs) at multiple different scales and using boosted decision trees for classification. Though resulting in higher accuracy, the higher dimensionality ACFs exacerbate memory operations, which become the energy and speed bottlenecks of the system. To overcome this, we employ binary discrete cosine transform to perform low-overhead and lossy compression, to efficiently store and access feature data. For restoring performance following compression, we exploit retraining of the classifier, resulting in an optimal model for pedestrian detection. The proposed accelerator is implemented in field-programmable gate array, which can process 40 video graphics array frames (640 × 480 resolution) per second at a log-average miss rate of 42% on the Caltech Pedestrian dataset, with compression reducing memory energy by 4× and overall energy by 1.7×.