Embedded video systems have restricted hardware resources and are required to work in real-time. Existing object detection architectures are far from reaching acceptable small object detection accuracy and efficiency for embedded devices. This thesis is an attempt to address the aforementioned issue by presenting a temporal-enhanced lightweight YOLO architecture which is highly energy efficient for small object detection in embedded video systems. Temporal context from adjacent frames is utilized in the suggested framework to improve consistency and accuracy in detection, and the architectural elements of the framework are lightweight, using depthwise separable convolutions and optimized feature fusion operations to decrease the computational load. The model is trained and tested on a real-world surveillance dataset with ten object classes such as cars, pedestrians, and bicycles. The experimental results confirm that significant progress over the baseline can be made on all the metrics. The results achieve 22.6 mean Average Precision at IoU 0.5 (mAP@0.5), which is a 44% absolute increase in detection performance, while also obtaining 23.6 frames per second (FPS) and 42.4 milliseconds (ms) mean latency, which is a 201% relative improvement on the frames per second per GFLOP (FPS/G) efficiency metric. Additionally, the network has 6.3 megabytes (MB) size with 3.16 million (M) parameters, which satisfies all the design constraints of being capable of real-time, lowlatency, and low-memory deployment on embedded devices. We show that our work benefits all classes in particular, the ones with small objects, which are especially difficult to detect and for which we achieve a significant speedup. We also provide a set of methods applicable to real-time computer vision systems on embedded devices, such as surveillance cameras, self-driving cars, or Internet of Things sensors, for which it is often considered that accuracy and speed cannot be jointly optimized in embedded video processing.
