Faster Object Detection in Videos for Autonomous Vehicle Applications

Autonomous vehicles (AVs) face the challenge of achieving high-quality object detection in resource-constrained environments. Our work focuses on enhancing the speed of state-of-the-art object detectors for AV-based applications without incurring substantial computational costs. Current approaches, like pruning and quantization, achieve lower accuracy than deep-learning models or require the development of new model architectures. We propose a novel preprocessing-cum-postprocessing algorithm, ConcatDet, which can be applied to pre-trained object detectors to improve their speed while maintaining similar accuracy. In preprocessing, we employ a concatenation method that allows simultaneous processing of two video frames by the object detector. In postprocessing, we apply non-maximum suppression to accommodate the temporal correlation between objects in the two frames, resulting in final predictions. ConcatDet achieves nearly 50% reduction in the inference time per frame compared to original models while incurring a 2-5% loss mAP on KITTI-MOT and Highway Driving datasets.