(I) Objective: Detect people, motorcycles, and cars in a nearly 180-degree rear panoramic view to enhance vehicle safety during reversing. 

(II) Approach: YOLOv7-tiny was chosen for its high frames per second (FPS) performance on the company’s low-compute-power chip, allowing it to capture high-speed vehicles approaching from behind with accuracy.  

(III) Dataset: Trained on open public datasets like BDD, Mapillary, NuScenes, IDD, and Waymo. Various augmentations, including rotation, translation, scaling, shearing, perspective transformation, and mosaic, were applied to help the model predict people, motorcycles, and cars accurately, even in distorted images.

(IV) Result: The YOLOv7-tiny model successfully detected people, motorcycles, and cars within a 10-15m range at a resolution of 480x256 pixels in the customer’s input image. 

The below images show the YOLOv7-tiny model's predictions, including person bounding boxes (blue), motorcycle bounding boxes (pink), and car bounding boxes (light blue).