Detecto is a learning-based object detection FPGA IP core, developed for embedded vision applications. The algorithm follows a discriminative approach: it combines a HOG-based descriptor and a SVM classifier. HOG (Histogram of Oriented Gradients) is a descriptor encoding the object structure. SVM (Support Vector Machine) is a non probabilistic binary linear classifier. The core support multiple scales to detect object moving in an arbitrary range of distance. The core is provided with a built-in pedestrian detection classifier trained on a wide range of automotive scenarios. User-defined classifiers can be loaded via software API.
Examples of object models are pedestrians, vehicles, traffic signs, animals, etc. Multiple SVM blocks can be instantiated to detect different objects in parallel. eVS provides software tools and technical support for training classifiers on specific customer application.
A detection window slides the input image. For each window position the HOG signature is computed and classified to establish if it is the target object or not.
Post-processing and distance estimation
The core is provided with drivers and an embedded software library in C code implementing post-processing functions necessary to group together overlapping detections, track the objects and compute their distance from the camera. The detection range depends on the imager and camera lens: with 1Mp @30fps camera, detection approximately covers a range of 1-10m using 180º FOV or 2-40m using a 50º FOV.
Detecto can be evaluated within the logiADAK Automotive Driver Assistance kit.
For any technical details or information about evaluation and licensing you can refer to our technology partner Xylon d.o.o. that is authorized to supply Detecto (named logiHOG) under the logicBRICKS™ brand. Click here to learn about Xylon's solution for the Pedestrian Detection.